Home Creators Posts Import Register Favorites Logout

Downloads

Content

Wan 2.1 Gradio app for text to video, image to video and video to video that runs with as low as 4 GB VRAM and batch processes whole folders.

Ultra advanced app with 1-click installers for Windows, RunPod, Massed Compute and SimplePod, batch processing, Torch 2.8, CUDA 12.9, Flash Attention, Sage Attention and xFormers, and support for older GPUs like the RTX 2000 series. Includes a full video tutorial, a Squishing LoRA tutorial and downloadable model zips.

1-Click to install and use the most powerful ever released Image to Video, Text to Video and Video to Video Wan 2.1 Open source model

Patreon exclusive posts index to find our scripts easily, Patreon scripts updates history to see which updates arrived to which scripts and amazing Patreon special generative scripts list that you can use in any of your task.

Join discord to get help, chat, discuss and also tell me your discord username to get your special rank : SECourses Discord

Please also Star, Watch and Fork our Stable Diffusion & Generative AI  GitHub repository and join our Reddit subreddit and follow me on LinkedIn (my real profile)

=======

Step by step full Video Tutorial : https://www.youtube.com/watch?v=hnAhveNy-8s

Squishing LoRA Tutorial (Tutorial number 2) : https://youtu.be/ueMrzmbdWBg

Latest installer : Wan21_v75.zip

Windows Requirements

3 November 2026 v74

Support for older GPUs implemented like RTX 2000 series

25 October 2025 v72

  • Libraries and installers upgraded to Torch 2.8, CUDA 12.9, latest Flash Attention, Sage Attention, xFormers all compiled by me for RTX 1000 series to RTX 5000 series

  • Prompt enhance bug fixed

2 June 2025

  • In some cases even 8 steps yields perfect results such as 480p image to video model and 512x768 px

  • This is discovered while I was testing CausVid LoRA

  • Sadly CausVid LoRA not working still, previous announcement was a mistake

28 April 2025 Update V68

  • Batch processing was not saving files with same name, this bug fixed

  • Now you can give custom outputs folder the format as below

    • python App.py --outputs "C:\fantasy_talking_tutorial\chars\a"

    • edit the Start_APP.bat file from latest zip file you will see

18 April 2025 Update V66

  • Now when you run multiple instances of the app on the same platform, they won't overwrite each other when saving the videos

  • Now batch processing will process webp images as well

  • Please run installer file to update requirements

13 April 2025 Update V64

  • Video to video improved and became more robust

  • Lets say you have uploaded 13 seconds 30 FPS video and set target frame count 81, it will get 81 - 1 / 16 = first 5 seconds of video and use it

    • It will re-encode first 5 seconds into 16 FPS since model supports only 16 FPS

  • If you want your output lets say 3 seconds then set 3 * 16 + 1 = 49 frames

  • It will also now transfer audio if your input video has audio

  • It will more robustly rename input videos to prevent errors

  • Auto pre-processed videos will be saved in auto_pre_processed_videos folder

  • Just run Windows_Update.bat to update from v60

  • Hopefully GGUF support and AI video upscale coming soon, sorry for delay

31 March 2025 Update V60

  • I have rented a 1000 GB - 64 Cores Linux server to compile Flash Attention and Sage Attention on Linux

    • Still it took huge time

  • Now RunPod and Massed Compute installers will use these pre-compiled wheels and they are 10x faster than before to install

    • Now RunPod and Massed Compute will use Sage Attention as default

    • Now RunPod and Massed Compute scripts supports RTX 5000 series GPUs as well

  • If you are Linux user, you can use RunPod installer script

  • Windows start file now will set Gradio temp folder as set TEMP=temp, set TMP=tmp, so Gradio generated temp files will be inside your local installation folder

  • With Sage Attention now L40 works really faster on Massed Compute

    • Select 48 GB config and apply fast preset for fastest results

26 March 2025 Update V57

  • Pipelines were getting reloaded totally unnecessarily even though none of the critical parameters changed

  • Now your next generation should immediately start if no model changes happened

  • Only following parameters changes will reload pipeline now - report otherwise

    • Used Model

    • LoRA Model

    • LoRA Scale

    • Number of Persistent Parameters In Dit (VRAM)

  • 16:9_low or 9:16_low aspect ratios were not saved in configs, now will work

  • Default Sigma Shift is now 5.6 and default CFG Scale is now 5

  • This change is made to reduce color changes especially when extending videos

  • Fast preset will also now set Sigma Shift as 5.6

  • Sadly the pipeline RAM management is not great, so too many pipeline deload and reload may cause application to crash if your RAM is not sufficient - I am trying my best to improve RAM management

  • Running Windows_Update.bat should be sufficient

25 March 2025 Update V56

  • The installer will now install and use Sage Attention instead of Flash Attention

    • Please download and use latest installer zip file

  • It is 37% faster at 20 steps, 81 frames, 1280x720p text to video model with Teacache 0.15

  • You can see a comparison video here : https://www.reddit.com/r/SECourses/comments/1jjgrl5/secourses_wan_21_ultra_advanced_app_is_now_37/

  • To have it, use latest zip file, extract and run installer

    • This will uninstall Flash Attention and install Sage Attention

  • For those who wants to use Flash Attention instead of Sage Attention, run Use_Flash_Attention_Instead_Of_Sage_Attention.bat after installing the App

    • This will uninstall Sage Attention and re-install Flash Attention

24 March 2025 Update V52

  • Windows_Update.bat should be sufficient to work

    • If you get errors make a fresh install and move model files

  • Now when changing models or GPU VRAM presets, selected aspect ratio should stay same

  • Batch processing logic completely re-written

  • Now both video extending and number of generations should work perfect in single generations and batch processing given folder (it can be image to video, last frame of videos, video to video)

  • New logic of file saving will be like below - e.g. extension count is set as 3 and RIFE enabled

    • 0001.mp4 - first original

    • 0001_ext1.mp4 - extended - 1st extension - uses same file name as original 1 - uses last frame of 1.mp4

    • 0001_ext2.mp4 - extended - 2nd extension - uses last frame of 1_ext1.mp4

    • 0001_improved.mp4 - first original improved with RIFE

    • 0001_ext1_improved.mp4 - 1st extension improved with RIFE

    • 0001_ext2_improved.mp4 - 2nd extension improved with RIFE

    • 0001_extended_original.mp4 - first original + 1st extension + 2nd extension - no RIFE yet

    • 0001_extended_original_enhanced.mp4 - first original + 1st extension + 2nd extension with RIFE

  • Please extensively test every feature and let me know

  • There are still somethings i need to fix and add some amazing new features hopefully tomorrow

23 March 2025 Update V51

  • Windows_Update.bat should be sufficient to work

    • If you get errors make a fresh install and move model files

  • Now prompt will be also saved and loaded with your configs

  • If you use multi‑line prompts, the way the seed is calculated has been adjusted. Now each line (or generation) gets a consistent seed value for a given generation round. This means that when you write several lines in your prompt, they are processed more uniformly.

  • Moreover, extension process was not properly working with Batch processing, these errors fixed

  • When you use the “Extend Video Factor” feature (to lengthen your videos by automatically generating additional segments), the process is now a bit smoother. The app now handles the extraction of the last frame and the re‐generation of extra segments with improved consistency

  • In batch mode too, if you are processing multiple files at once, the extension functionality has been improved and bugs fixed

  • Unnecessary pipeline reloading fixed to improve processing speed

  • The app now makes sure that when it generates extension segments, it correctly applies any LoRA models you were using in your initial video generation. This ensures a more consistent style and look throughout the extended video.

22 March 2025 Update V50

  • Interface organization improved, many more updates coming soon hopefully

  • Windows_Update.bat should be sufficient to work

    • If you get errors make a fresh install and move model files

  • There were several errors with RIFE FPS increase when using new Extend Video Factor feature (boomerang effect) and they are fixed

  • Now it will work with below workflow

    • Lets say you selected WAN 2.1 1.3B (Text/Video-to-Video) as a beginning

    • Set Extend Video Factor = 3

    • It will generate 1st video with WAN 2.1 1.3B and save it : 0001.mp4

      • If RIFE enabled, it will also save improved_0001.mp4

      • If Save prompt to file enabled it will save 0001.txt

    • Then it will get last frame of 0001.mp4 and use WAN 2.1 14B Image-to-Video 480P and generate the second video : will save as 0002.mp4

      • If RIFE enabled, it will also save improved_0002.mp4

      • If Save prompt to file enabled it will save 0002.txt

      • This txt file will show which video's last frame used

    • Then it will get last frame of 0002.mp4 and use WAN 2.1 14B Image-to-Video 480P and generate the third video : will save as 0003.mp4

      • If RIFE enabled, it will also save improved_0003.mp4

      • If Save prompt to file enabled it will save 0003.txt

    • Then it will merge 0001.mp4 + 0002.mp4 + 0002.mp4 as 00001_extended_original.mp4

    • If RIFE, then it will merge improved_0001.mp4 + improved_0002.mp4 + improved_0003.mp4 as 00001_extended_enhanced.mp4

  • When user sets static seed and uses multi-line prompting, it was incrementing seed by 1

    • This erroneous behavior fixed

    • When you use more than 1 generation feature, number of generations, they will continue to increase seed by 1 as expected and works with multi-line as well

Older Update History : https://github.com/FurkanGozukara/Stable-Diffusion/blob/main/App_Updates_Change_Logs/Wan2.1_App.md

How To Install and Use:

Check above attached screenshots very carefully

Windows:

  • First watch above requirements tutorial and follow every step

    • Extract zip archive into a drive like c:/wan_gen

    • Longer directory path causes issues

    • Double click and install with Windows_Install.bat

    • The double click Windows_Download_Models.bat and download models you want

    • Then run with Windows_Start_App.bat

Massed Compute (Recommend Cloud) :

RunPod (Cloud):

Related posts

All posts in this topic: AI Video Generation and Animation Apps · New to SECourses? Start Here: Best and Most Popular SECourses Posts

Files

Comments

Ec Jep

Thx for all the work. I’ll try it soon. Can you use any custom models or only WAN 2.1.1?

FredrickFlower

I'm getting a paths don't have the same drive error when trying to download the models?

SECourses

never seen it before. please give me full logs.

SECourses

this published today there isnt any custom model as far as i know :D

SECourses

yes works with python 3.10. if you use 3.11 or 3.12 you need to change some of the pre installed wheels

BecauseReasons

Does not work. Have tried to install twice. CUDA cards unavailable error; even though CUDA IS available.

SECourses

Fixed error use v2 zip start bat file pls

FredrickFlower

Enter your choice (1-4): 2 Downloading Wan2.1-I2V-14B-720P... Traceback (most recent call last): File "O:\Wan2.1\Download_Models.py", line 51, in download_model(sys.argv[1]) File "O:\Wan2.1\Download_Models.py", line 19, in download_model snapshot_download( File "O:\Python\lib\site-packages\huggingface_hub\utils\_validators.py", line 120, in _inner_fn return fn(*args, **kwargs) File "O:\Python\lib\site-packages\huggingface_hub\_snapshot_download.py", line 233, in snapshot_download _inner_hf_hub_download(file) File "O:\Python\lib\site-packages\huggingface_hub\_snapshot_download.py", line 212, in _inner_hf_hub_download return hf_hub_download( File "O:\Python\lib\site-packages\huggingface_hub\utils\_validators.py", line 120, in _inner_fn return fn(*args, **kwargs) File "O:\Python\lib\site-packages\huggingface_hub\file_download.py", line 1280, in hf_hub_download return _to_local_dir(blob_path, local_dir, relative_filename, use_symlinks=local_dir_use_symlinks) File "O:\Python\lib\site-packages\huggingface_hub\file_download.py", line 1568, in _to_local_dir _create_symlink(real_blob_path, local_dir_filepath, new_blob=False) File "O:\Python\lib\site-packages\huggingface_hub\file_download.py", line 873, in _create_symlink _support_symlinks = are_symlinks_supported(os.path.dirname(os.path.commonpath([abs_src, abs_dst]))) File "O:\Python\lib\ntpath.py", line 807, in commonpath raise ValueError("Paths don't have the same drive") ValueError: Paths don't have the same drive Press any key to continue . . .

Alex

When I click on Generate I get an error with FlashAttention "RuntimeError: FlashAttention only supports Ampere GPUs or newer" My graphics card is old (NVIDIA GeForce RTX 2060 with 6GB of VRAM). Is there a solution or not?

Arcadius

it doesn't work, all it does it auto close the window batch scripts

Alex

maybe use another attention mode like sdpa ?

EvilGiggles

How much RAM does the image to video model use? When I press Generate, it maxes out my 64GB RAM and takes a very long time to start. And once it starts, I'm getting 100+sec / it... (Using the 24GB VRAM preset with a 4090).

hacbachvotinh

i think we must have 128gb System Ram to use 14b model.

hacbachvotinh

i got this error when use video2video: File "e:\wan21_v2\wan2.1\diffsynth-studio\diffsynth\pipelines\wan_video.py", line 224, in __call__ latents = self.scheduler.add_noise(latents, noise, timestep=self.scheduler.timesteps[0]) File "e:\wan21_v2\wan2.1\diffsynth-studio\diffsynth\schedulers\flow_match.py", line 65, in add_noise sample = (1 - sigma) * original_samples + sigma * noise RuntimeError: The size of tensor a (113) must match the size of tensor b (21) at non-singleton dimension 2

cool1

Thanks a lot of doing this video model. Though the model files it downloads are .pth files instead of .safetensors (other video models like cosmos and skyreels use the safe .safetensors file format). perplexity.ai says ".pth files can contain and execute arbitrary Python code, making them a potential vector for malicious attacks". Has anyone done a check on the .pth model files to make sure they're safe (no pickle stuff?)? edit: can the model files be converted to .safetensors with something like this: https://github.com/MackinationsAi/convert-2-safetensors and if they can will the app files then need to be edited to load the new format? It should be a lot less possible risk if it can work with them converted to .safetensors I think. Also some files it downloads are .pth but the bigger model files it downloads seems to be .safetensors. The vae is .pth.

SECourses

well it downloads from whatever they published :) i am pretty sure safe

SECourses

yes make your video sizes and image sizes matching the resolutions. that is why. i will try to add auto crop today

SECourses

it works on my 96gb ram and 24 gb gpu but really slow :(

SECourses

yes 100+ second normal. it is 30 second even on h200 :D

SECourses

you need to follow requirements video first : https://youtu.be/DrhUHnYfwC0 you probably missing python. also you can start a cmd and type bat file name and it wont auto close

SECourses

sadly i dont know how can it be converted atm. yes this utilizes flash attention i installed pre compiled wheel. i will look for a solution

cool1

Thanks. Though we don't know how safe they are unless someone's checked them. Converting to .safetensors for the file(s) that are that format (like vae) should be safest. Maybe I or someone can convert the .pth one and then edit the .py file to load the right one if it's a one line edit of the filename - or maybe it would be harder than that.

SECourses

well i checked with kaspersky no issues. by the way this is a global company with billion dollars i dont think they will add anything into them :D

SECourses

is this O your physical drive? can you try on C drive? this looks like error of huggingface_hub library can you edit the bat file and remove set HF_HUB_ENABLE_HF_TRANSFER=1 and try again and let me know?

SECourses

I just asked this to the authors for older gpus

Alex

Thanks! Yeah, because when I use Hunyuan in ComfyUI, for example, I select SDPA instead of Flash, since Flash Attention doesn't work on older cards.

puk

THX for your work, i get this error: WARNING: Failed to find CUDA. But the app starts. img2Video needs over 1 hour

SECourses

hi please download v2. v1 zip start file has set cuda visible devices that is why it fails to find cuda. also v3 coming hopefully soon with more features

puk

THX where i can download the v2. v1?

SECourses

wait v3 almost ready . it will have more amazing features

puk

THX, made a fresh install but i get the same error: WARNING: Failed to find CUDA INFORMATION: No files could be found with the specified patterns can be found. (translated from german 2 english)

puk

RTX 4090

SECourses

yes it should work. so make a fresh install and send me install logs : monstermmorpg@gmail.com you probably missing requirements. if you upgrade gold member i can connect your pc and install for you

Pak26

I get an error when trying to install models: (…)pytorch_model-00001-of-00007.safetensors: 9%|██▉ | 870M/9.85G [03:15<33:32, 4.46MB/s] Traceback (most recent call last): File "C:\Users\pbour\AppData\Local\Programs\Python\Python310\lib\site-packages\huggingface_hub\file_download.py", line 498, in http_get hf_transfer.download( Exception: Error while removing corrupted file: The process cannot access the file because it is being used by another process. (os error 32) The above exception was the direct cause of the following exception: Traceback (most recent call last): File "C:\wan_gen\Download_Models.py", line 51, in download_model(sys.argv[1]) File "C:\wan_gen\Download_Models.py", line 19, in download_model snapshot_download( File "C:\Users\pbour\AppData\Local\Programs\Python\Python310\lib\site-packages\huggingface_hub\utils\_validators.py", line 118, in _inner_fn return fn(*args, **kwargs) File "C:\Users\pbour\AppData\Local\Programs\Python\Python310\lib\site-packages\huggingface_hub\_snapshot_download.py", line 306, in snapshot_download _inner_hf_hub_download(file) File "C:\Users\pbour\AppData\Local\Programs\Python\Python310\lib\site-packages\huggingface_hub\_snapshot_download.py", line 283, in _inner_hf_hub_download return hf_hub_download( File "C:\Users\pbour\AppData\Local\Programs\Python\Python310\lib\site-packages\huggingface_hub\utils\_validators.py", line 118, in _inner_fn return fn(*args, **kwargs) File "C:\Users\pbour\AppData\Local\Programs\Python\Python310\lib\site-packages\huggingface_hub\file_download.py", line 1457, in hf_hub_download http_get( File "C:\Users\pbour\AppData\Local\Programs\Python\Python310\lib\site-packages\huggingface_hub\file_download.py", line 509, in http_get raise RuntimeError( RuntimeError: An error occurred while downloading using `hf_transfer`. Consider disabling HF_HUB_ENABLE_HF_TRANSFER for better error handling. Press any key to continue . . . Thanks

SECourses

yes edit the download models bat file with notepad and remove set HF_HUB_ENABLE_HF_TRANSFER=1 you can use any notepad editor can you also reply here your system info we are trying to get this bug fixed : https://github.com/huggingface/hf_transfer/issues/63#issuecomment-2685182565 if you tell me information i can also reply Location and all network information (WiFi, cable, mobile, ideally provider). The disk that is written on (SSD, HDD, network mounted..)

Pak26

I disabled my VPN (NordVPN) and it seems to work now ! Still downloading... Enter your choice (1-4): 2 Downloading Wan2.1-I2V-14B-720P... C:\Users\pbour\AppData\Local\Programs\Python\Python310\lib\site-packages\huggingface_hub\file_download.py:149: UserWarning: `huggingface_hub` cache-system uses symlinks by default to efficiently store duplicated files but your machine does not support them in C:\. Caching files will still work but in a degraded version that might require more space on your disk. This warning can be disabled by setting the `HF_HUB_DISABLE_SYMLINKS_WARNING` environment variable. For more details, see https://huggingface.co/docs/huggingface_hub/how-to-cache#limitations. To support symlinks on Windows, you either need to activate Developer Mode or to run Python as an administrator. In order to see activate developer mode, see this article: https://docs.microsoft.com/en-us/windows/apps/get-started/enable-your-device-for-development warnings.warn(message) (…)pytorch_model-00001-of-00007.safetensors: 100%|███████████████████████████████▉| 9.85G/9.85G [19:27<00:00, 8.44MB/s] (…)pytorch_model-00002-of-00007.safetensors: 12%|███▊ | 1.15G/9.80G [03:18<15:54, 9.05MB/s] Thanks

Pak26

Do you think it's possible to run the i2V model with a rtx 3060 12Gb ?

SECourses

at lower resolution could be. 1280x720 takes 3 hours on rtx 3090 ti for 50 steps

Ans Ass

48gb vram minimum? On 24gb it overflows into RAM for over 10gb making it extremely slow and unusable.

Flambo

Converting them would still run the same risk as the same executable code sections are still loaded during conversion, granted you could do this in an isolated environments but yeah..

cool1

I wondered that too. The code is there in that github to check if needed. I thought the difference might be that the conversion wasn't designed to execute any code whereas running the model normally would. I asked deepseek r1 14b to check the converter code and these are some things it said: Loading Mechanism: The script uses torch.load with map_location="cpu", ensuring that any loaded tensors are processed on CPU, avoiding GPU execution which could be riskier. Code Execution Risks: While .pth files can theoretically execute code upon loading, this script specifically targets state_dict conversion, minimizing the risk of arbitrary code execution beyond tensor operations. .. Dependency Risks: Relies on PyTorch, which has a history of being secure for such operations, though vulnerabilities are always possible. ...Conclusion: The script appears designed to safely convert .pth files by focusing solely on tensor data and including verification steps. However, as with any external code, it's advisable to run it in a controlled environment and verify its functionality before use. -- Though each file would only need to be converted once by someone who had the right sandboxed environment. I thought there was an online converter that might be less risky (though might need a high upload bandwidth if they're high file sizes) though one website/perplexity is saying that might not be totally safe either (an article mentions potential hijacking of that online conversion I think).

SECourses

14b models are working at low vram but slow. 1.3b model works as low as 3.5 gb

Ans Ass

Thanks for your reply, and also for the 14-hour shift to get this out asap! 48g seems to be the most likely option for image-2-video. What was your experience on 32gb 5090 for i2v?

Ahmet Inceelli

Hello, Thanks for your 14 hours effort for us. 😸 I installed on Massed Compute Text to Video 1.3B is OK But the process suddenly stops on these models below Text to Video 14B and Image to Video 14B Ubuntu@0030-ict-prxmx50125:~/apps/Wan21_v2$ cd Wan2.1 export PYTHONWARNINGS="ignore" source venv/bin/activate export HF_HUB_ENABLE_HF_TRANSFER=1 python3 App.py --share * Running on local URL: http://127.0.0.1:7860 * Running on public URL: https://082ceb5ff71143e235.gradio.live This share link expires in 72 hours. For free permanent hosting and GPU upgrades, run `gradio deploy` from the terminal in the working directory to deploy to Hugging Face Spaces (https://huggingface.co/spaces) [CMD] Loading model: 1.3B with torch dtype: torch.bfloat16 and num_persistent_param_in_dit: 12000000000 Loading models from: models/Wan-AI/Wan2.1-T2V-1.3B/diffusion_pytorch_model.safetensors model_name: wan_video_dit model_class: WanModel This model is initialized with extra kwargs: {'model_type': 't2v', 'patch_size': (1, 2, 2), 'text_len': 512, 'in_dim': 16, 'dim': 1536, 'ffn_dim': 8960, 'freq_dim': 256, 'text_dim': 4096, 'out_dim': 16, 'num_heads': 12, 'num_layers': 30, 'window_size': (-1, -1), 'qk_norm': True, 'cross_attn_norm': True, 'eps': 1e-06} The following models are loaded: ['wan_video_dit']. Loading models from: models/Wan-AI/Wan2.1-T2V-1.3B/models_t5_umt5-xxl-enc-bf16.pth model_name: wan_video_text_encoder model_class: WanTextEncoder The following models are loaded: ['wan_video_text_encoder']. Loading models from: models/Wan-AI/Wan2.1-T2V-1.3B/Wan2.1_VAE.pth model_name: wan_video_vae model_class: WanVideoVAE The following models are loaded: ['wan_video_vae']. Using wan_video_text_encoder from models/Wan-AI/Wan2.1-T2V-1.3B/models_t5_umt5-xxl-enc-bf16.pth. Using wan_video_dit from models/Wan-AI/Wan2.1-T2V-1.3B/diffusion_pytorch_model.safetensors. Using wan_video_vae from models/Wan-AI/Wan2.1-T2V-1.3B/Wan2.1_VAE.pth. No wan_video_image_encoder models available. num_persistent_val 12000000000 [CMD] Model loaded successfully. [CMD] Generating video 1/1 with prompt: Young girls play volleyball in the gym. width: 832 : height: 480 100%|███████████████████████████████████████████| 50/50 [16:03<00:00, 19.27s/it] VAE decoding: 100%|█████████████████████████████| 18/18 [01:42<00:00, 5.69s/it] Saving video: 100%|██████████████████████████| 161/161 [00:01<00:00, 104.34it/s] [CMD] Saved video: outputs/00002.mp4 [CMD] Saved prompt: outputs/00002.txt [CMD] Generation complete. Duration: 1114.46 seconds. Last used seed: 3178083971 WAN 2.1 14B Text-to-Video [CMD] Loading model: 14B_text with torch dtype: torch.bfloat16 and num_persistent_param_in_dit: 22000000000 Loading models from: ['models/Wan-AI/Wan2.1-T2V-14B/diffusion_pytorch_model-00001-of-00006.safetensors', 'models/Wan-AI/Wan2.1-T2V-14B/diffusion_pytorch_model-00002-of-00006.safetensors', 'models/Wan-AI/Wan2.1-T2V-14B/diffusion_pytorch_model-00003-of-00006.safetensors', 'models/Wan-AI/Wan2.1-T2V-14B/diffusion_pytorch_model-00004-of-00006.safetensors', 'models/Wan-AI/Wan2.1-T2V-14B/diffusion_pytorch_model-00005-of-00006.safetensors', 'models/Wan-AI/Wan2.1-T2V-14B/diffusion_pytorch_model-00006-of-00006.safetensors'] model_name: wan_video_dit model_class: WanModel This model is initialized with extra kwargs: {'model_type': 't2v', 'patch_size': (1, 2, 2), 'text_len': 512, 'in_dim': 16, 'dim': 5120, 'ffn_dim': 13824, 'freq_dim': 256, 'text_dim': 4096, 'out_dim': 16, 'num_heads': 40, 'num_layers': 40, 'window_size': (-1, -1), 'qk_norm': True, 'cross_attn_norm': True, 'eps': 1e-06} Killed

ApathyJones

I had the same issue. I fixed it by un-commenting the REM SET CUDA_VISIBLE_DEVICES=1 (by removing the REM in front) and setting it to 0 instead of 1.

Pak26

On the Huggingface of the model it says there's a I2V-14B-480P that support 480p. Maybe it could be a better choice to run this model with my rtx 3060 12Gb ? Do we have the option I2V-14B-480P in the model installer? Thanks

The90sPope

Struggle with Visual studio, 17.13 wont let install CUDA, and 17.14 preview apparently is broken too. Have no ways of getting CUDA t o work. EDIT: got 2022 visual studio 1714 preview to work with CUDA installs. DO NOT USE 17.13 VISUAL STUDIO just a heads up for all the fresh people like me. It will get your CUDA installation stuck.

SECourses

out of RAM simply. so get 2x GPU machine or stronger machine :D

SECourses

around 70 second / it. even h200 had like 30 second / it

SECourses

it is still 14b so no difference :D you can use our config at lower resolution i think should work fairly well

SECourses

yes that one can prevent download of models

The90sPope

I did follow, if I use it, CUDA will not find Visual Studio at all. Tried it. But 17.14 preview appears to work. I am on win11

Brian Rivera

Hey, I got an error when attempting to generate (RuntimeError: FlashAttention only supports Ampere GPUs or newer.). My GPU is a 2070, so I know it is Turing and not Ampere. I was just curious if you knew of any workarounds, though my hopes are not high. Thanks!

SECourses

sadly not yet. we are looking for a solution

SECourses

if working great. i installed all these versions with this very weird

Ahmet Inceelli

L40 x 1 is OK? Other GPUs are not available?

SECourses

L40 should work. let me update v6 it will ensure it works with virtual ram

SECourses

I just added 480p image to video model took less than 20 min on rtx 5090 50 steps highest quality

Pak26

I tried to run the app but when i click Genarate the app seems to load the model but then it disconnect "Error connection error out"? Any idea what is the problem ?

The90sPope

Generating first video on 1.3b with 16gb vram, I have 4080super. It does allocate and use 16gb vram with 16gb preset, but gpu core remains at 0%, so it look like only memory is utilized. It has been over 20mins and still going.

The90sPope

It appears that my gpu cores are not utilized while generating on 1.3b model on rtx4080super. Using 16gb preset, was generating over 35minutes, so I gave up. Something did not work I assume. It should not take this long on 16gb vram

The90sPope

I found the issue, it was the prompt enhance, without it, i used under 7 minutes for 50 steps on 1.3b

Michał Kowalski

First of all, thank you for another great project. Installing that app like the rest 13 apps to that moment, finished without one error, no cuda, python problems and visual. I don't have to switch env path, just did tutorial step by step word by word as you sad :) and today I am sure that second gpu 5090 will be very helpfull in future;). Actually I have 4090, 64RAM DDR5,i9, Downloaded all models, After ran the app I checked 16GB, chose image to video 14b, ratio 16:9 uploaded one input image 842 x 652 Pixel, image to video, insert prompt: there is earthquake and all residences are falling down into precipice and didn't change any other defaults. It is running without issues, Ram is full in 80% but it is stable, GPU 95-100% (about, stable too with funs on 80%),i9 doesn;t feel it at all, actually I wanted to stop it because of about 3 hours of waiting but I see now that it getting the process faster and it shows now about 1,30h/160.15s/it 17/50 inference steps. I will wait for results and then test thinner models. Additionally guys I recommend special cooling for your fastest nvme like kingston fury, after removing labels from both sides and insert it to AXAGON case made whole from radiator with thermal pads and active fan plugged directly to mainboard have 45-50 celsius in high stress, it is matter for long actions with Ai and generating so much heat which I guess helped me to burn first Kingston ;) update :) I received 5s video it took: "Generating video 1/1 with prompt: there is earthquake and all residences are falling down into precipice" quite nice result for the begining width: 1280 : height: 720 100%|███████████████████████████████████████████████████████████████████████████████| 50/50 [2:06:13<00:00, 151.48s/it] VAE decoding: 100%|████████████████████████████████████████████████████████████████████| 30/30 [00:47<00:00, 1.60s/it] Saving video: 100%|███████████████████████████████████████████████████████████████████| 81/81 [00:00<00:00, 195.12it/s] [CMD] Saved video: outputs\00001.mp4 [CMD] Saved prompt: outputs\00001.txt [CMD] Generation complete. Duration: 7691.67 seconds. Last used seed: 3099509642 and tested 1.3 model [CMD] Loading model: 1.3B with torch dtype: torch.bfloat16 and num_persistent_param_in_dit: 7000000000 Loading models from: models/Wan-AI/Wan2.1-T2V-1.3B/diffusion_pytorch_model.safetensors model_name: wan_video_dit model_class: WanModel This model is initialized with extra kwargs: {'model_type': 't2v', 'patch_size': (1, 2, 2), 'text_len': 512, 'in_dim': 16, 'dim': 1536, 'ffn_dim': 8960, 'freq_dim': 256, 'text_dim': 4096, 'out_dim': 16, 'num_heads': 12, 'num_layers': 30, 'window_size': (-1, -1), 'qk_norm': True, 'cross_attn_norm': True, 'eps': 1e-06} The following models are loaded: ['wan_video_dit']. Loading models from: models/Wan-AI/Wan2.1-T2V-1.3B/models_t5_umt5-xxl-enc-bf16.pth model_name: wan_video_text_encoder model_class: WanTextEncoder The following models are loaded: ['wan_video_text_encoder']. Loading models from: models/Wan-AI/Wan2.1-T2V-1.3B/Wan2.1_VAE.pth model_name: wan_video_vae model_class: WanVideoVAE The following models are loaded: ['wan_video_vae']. Using wan_video_text_encoder from models/Wan-AI/Wan2.1-T2V-1.3B/models_t5_umt5-xxl-enc-bf16.pth. Using wan_video_dit from models/Wan-AI/Wan2.1-T2V-1.3B/diffusion_pytorch_model.safetensors. Using wan_video_vae from models/Wan-AI/Wan2.1-T2V-1.3B/Wan2.1_VAE.pth. No wan_video_image_encoder models available. num_persistent_val 7000000000 [CMD] Model loaded successfully. [CMD] Generating video 1/1 with prompt: there is earthquake and all residences are falling down into precipice and fire, there is nothing left but smoke width: 832 : height: 480 VAE encoding: 100%|██████████████████████████████████████████████████████████████████████| 9/9 [00:09<00:00, 1.01s/it] 100%|██████████████████████████████████████████████████████████████████████████████████| 50/50 [04:11<00:00, 5.02s/it] VAE decoding: 100%|██████████████████████████████████████████████████████████████████████| 9/9 [00:14<00:00, 1.63s/it] Saving video: 100%|███████████████████████████████████████████████████████████████████| 81/81 [00:00<00:00, 843.53it/s] [CMD] Saved video: outputs\00002.mp4 [CMD] Saved prompt: outputs\00002.txt [CMD] Generation complete. Duration: 293.83 seconds. Last used seed: 4077048855 well it was quite faster based on created movie, results are ok, everything works perfect, only test it on different configs :)

lemas espandon

i keep getting this error: "RuntimeError: The size of tensor a (6) must match the size of tensor b (3) at non-singleton dimension 2" while trying video2video, the input video is the same resolution as the output, 832x480

Draxel

Fantastic addition to your Patreon. Thank you!

Michał Kowalski

Hi, what type of ram do you have? I have DDR5 64, chose 16GB for 4090 and it took 80% of RAM but almost 95-100% GPU working stable it will take about 2h I guess for 14b model

DanO..

I'm very excited about this. Thanks for the effort. First off you may want to warn people that loading all the models you end up just shy of 170GBs. Does the prompt enhancer need to be such a behemoth? 11.8GB of 12GB VRAM, 16.9 of 20GB shared VRAM. And when it's done it doesn't release it. Starting video generation does release the memory though. And, most importantly, 12GB RTX3060. Tried 14B t2v at default settings... my computer took some notes and restarted. So I tried image to video but at 320/368pixels. Same thing. Computer blue screen informing me it's restarting but taking notes first. I'm guessing something is wrong. It shouldn't just restart my machine if it runs out of RAM. Thanks again! -1.3B worked fine. Pretty ugly though! :)

puk

I made a new CUDA install. That fixed my Problem.

SECourses

Thanks a lot for huge info. Hard to know every cases to cover :D

SECourses

some people getting but i dont know how yet. my example videos working. i presume related to video codec. can you try to reencode and try again? if it is a safe video can you email me? monstermmorpg@gmail.com maybe i can figure out and solve

SECourses

nice. i am also recording a tutorial video showing timings etc :D

SECourses

you sure you selected 1.3b model? it uses max 6.5 GB vram and should be done in less than like 10 minutes

SECourses

ye prompt enhance uses a lot of VRAM :D

SECourses

what do you see on CMD? i also dded 480p model to the app

SECourses

awesome. 64 GB ram i recommend min.

lemas espandon

the video i'm using is nothing special just a random one i was trying to use as a test to see how it works, what codec is working for you in your examples?

SECourses

here example codec General Complete name : C:\Users\Furkan\Downloads\00010.mp4 Format : MPEG-4 Format profile : Base Media Codec ID : isom (isom/iso2/avc1/mp41) File size : 662 KiB Duration : 5 s 400 ms Overall bit rate : 1 004 kb/s Frame rate : 15.000 FPS Writing application : Lavf61.7.100 Video ID : 1 Format : AVC Format/Info : Advanced Video Codec Format profile : High@L3 Format settings : CABAC / 4 Ref Frames Format settings, CABAC : Yes Format settings, Reference fra : 4 frames Codec ID : avc1 Codec ID/Info : Advanced Video Coding Duration : 5 s 400 ms Bit rate : 1 001 kb/s Width : 480 pixels Height : 832 pixels Display aspect ratio : 0.577 Frame rate mode : Constant Frame rate : 15.000 FPS Color space : YUV Chroma subsampling : 4:2:0 Bit depth : 8 bits Scan type : Progressive Bits/(Pixel*Frame) : 0.167 Stream size : 660 KiB (100%) Writing library : x264 core 164 r3192 c24e06c Encoding settings : cabac=1 / ref=3 / deblock=1:0:0 / analyse=0x3:0x113 / me=hex / subme=7 / psy=1 / psy_rd=1.00:0.00 / mixed_ref=1 / me_range=16 / chroma_me=1 / trellis=1 / 8x8dct=1 / cqm=0 / deadzone=21,11 / fast_pskip=1 / chroma_qp_offset=-2 / threads=26 / lookahead_threads=4 / sliced_threads=0 / nr=0 / decimate=1 / interlaced=0 / bluray_compat=0 / constrained_intra=0 / bframes=3 / b_pyramid=2 / b_adapt=1 / b_bias=0 / direct=1 / weightb=1 / open_gop=0 / weightp=2 / keyint=250 / keyint_min=15 / scenecut=40 / intra_refresh=0 / rc_lookahead=40 / rc=crf / mbtree=1 / crf=25.0 / qcomp=0.60 / qpmin=0 / qpmax=69 / qpstep=4 / ip_ratio=1.40 / aq=1:1.00 Codec configuration box : avcC

Jeff Eberlin

Installs fine, when I try to generate with a 4090 using the 720p model it outputs this and just crashes: [CMD] Loading model: 14B_image_720p with torch dtype: torch.bfloat16 and num_persistent_param_in_dit: 0 Loading models from: ['models/Wan-AI/Wan2.1-I2V-14B-720P/diffusion_pytorch_model-00001-of-00007.safetensors', 'models/Wan-AI/Wan2.1-I2V-14B-720P/diffusion_pytorch_model-00002-of-00007.safetensors', 'models/Wan-AI/Wan2.1-I2V-14B-720P/diffusion_pytorch_model-00003-of-00007.safetensors', 'models/Wan-AI/Wan2.1-I2V-14B-720P/diffusion_pytorch_model-00004-of-00007.safetensors', 'models/Wan-AI/Wan2.1-I2V-14B-720P/diffusion_pytorch_model-00005-of-00007.safetensors', 'models/Wan-AI/Wan2.1-I2V-14B-720P/diffusion_pytorch_model-00006-of-00007.safetensors', 'models/Wan-AI/Wan2.1-I2V-14B-720P/diffusion_pytorch_model-00007-of-00007.safetensors'] Press any key to continue . . .

SECourses

did you download models? verify them. also how much RAM you have? did you test 1.3b model works?

lemas espandon

i can't figure it out, i don't even get an output from the auto-crop and if i make the video manually 16fps, 5s, 832x480, mpeg-4 it still does not work it does work if i use an output from the img2video

SECourses

if you can send me not working video i can try to debug

lemas espandon

I sent a pm on patreon with the video

Pak26

INFO: Could not find files for the given pattern(s). * Running on local URL: http://127.0.0.1:$$$$ To create a public link, set `share=True` in `launch()`. WAN 2.1 14B Image-to-Video [CMD] Loading model: 14B_image with torch dtype: torch.bfloat16 and num_persistent_param_in_dit: 12000000000 Loading models from: ['models/Wan-AI/Wan2.1-I2V-14B-720P/diffusion_pytorch_model-00001-of-00007.safetensors', 'models/Wan-AI/Wan2.1-I2V-14B-720P/diffusion_pytorch_model-00002-of-00007.safetensors', 'models/Wan-AI/Wan2.1-I2V-14B-720P/diffusion_pytorch_model-00003-of-00007.safetensors', 'models/Wan-AI/Wan2.1-I2V-14B-720P/diffusion_pytorch_model-00004-of-00007.safetensors', 'models/Wan-AI/Wan2.1-I2V-14B-720P/diffusion_pytorch_model-00005-of-00007.safetensors', 'models/Wan-AI/Wan2.1-I2V-14B-720P/diffusion_pytorch_model-00006-of-00007.safetensors', 'models/Wan-AI/Wan2.1-I2V-14B-720P/diffusion_pytorch_model-00007-of-00007.safetensors'] Press any key to continue . . . ...also the word "error" in the gui written in the "generated video" window The process just stop without any other message in the console... PS: I put $ to hide the rest of my address

juggz143

Why do you make 2 posts of the same thing? You should consider mentioning in the title what the difference is between the two, otherwise it is confusing to the user as to why they should use one versison over the other.

Wktra

I installed it at the root of my directory. Everything went well, but it won't start: `Traceback (most recent call last): File "Q:\Wan21_v7\Wan2.1\App.py", line 15, in import wan File "Q:\Wan21_v7\Wan2.1\wan\__init__.py", line 1, in from . import configs, distributed, modules File "Q:\Wan21_v7\Wan2.1\wan\modules\__init__.py", line 2, in from .model import WanModel File "Q:\Wan21_v7\Wan2.1\wan\modules\model.py", line 8, in from diffusers.models.modeling_utils import ModelMixin File "Q:\Wan21_v7\Wan2.1\venv\lib\site-packages\diffusers\models\modeling_utils.py", line 35, in from ..quantizers import DiffusersAutoQuantizer, DiffusersQuantizer File "Q:\Wan21_v7\Wan2.1\venv\lib\site-packages\diffusers\quantizers\__init__.py", line 15, in from .auto import DiffusersAutoQuantizer File "Q:\Wan21_v7\Wan2.1\venv\lib\site-packages\diffusers\quantizers\auto.py", line 22, in from .bitsandbytes import BnB4BitDiffusersQuantizer, BnB8BitDiffusersQuantizer File "Q:\Wan21_v7\Wan2.1\venv\lib\site-packages\diffusers\quantizers\bitsandbytes\__init__.py", line 2, in from .utils import dequantize_and_replace, dequantize_bnb_weight, replace_with_bnb_linear File "Q:\Wan21_v7\Wan2.1\venv\lib\site-packages\diffusers\quantizers\bitsandbytes\utils.py", line 35, in import accelerate File "Q:\Wan21_v7\Wan2.1\venv\lib\site-packages\accelerate\__init__.py", line 16, in from .accelerator import Accelerator File "Q:\Wan21_v7\Wan2.1\venv\lib\site-packages\accelerate\accelerator.py", line 36, in from accelerate.utils.imports import is_torchao_available File "Q:\Wan21_v7\Wan2.1\venv\lib\site-packages\accelerate\utils\__init__.py", line 14, in from .ao import convert_model_to_fp8_ao, filter_first_and_last_linear_layers, has_ao_layers File "Q:\Wan21_v7\Wan2.1\venv\lib\site-packages\accelerate\utils\ao.py", line 28, in from torchao.float8.float8_linear import Float8LinearConfig File "Q:\Wan21_v7\Wan2.1\venv\lib\site-packages\torchao\__init__.py", line 50, in from torchao.quantization import ( File "Q:\Wan21_v7\Wan2.1\venv\lib\site-packages\torchao\quantization\__init__.py", line 7, in from torchao.kernel import ( File "Q:\Wan21_v7\Wan2.1\venv\lib\site-packages\torchao\kernel\__init__.py", line 1, in from torchao.kernel.intmm import int_scaled_matmul, safe_int_mm File "Q:\Wan21_v7\Wan2.1\venv\lib\site-packages\torchao\kernel\intmm.py", line 10, in from torchao.kernel import intmm_triton File "Q:\Wan21_v7\Wan2.1\venv\lib\site-packages\torchao\kernel\intmm_triton.py", line 4, in import triton File "Q:\Wan21_v7\Wan2.1\venv\lib\site-packages\triton\__init__.py", line 8, in msvc_winsdk_inc_dirs, _ = find_msvc_winsdk() File "Q:\Wan21_v7\Wan2.1\venv\lib\site-packages\triton\windows_utils.py", line 230, in find_msvc_winsdk winsdk_inc_dirs, winsdk_lib_dirs = find_winsdk() File "Q:\Wan21_v7\Wan2.1\venv\lib\site-packages\triton\windows_utils.py", line 207, in find_winsdk winsdk_base_path, version = find_winsdk_registry() TypeError: cannot unpack non-iterable NoneType object Press any key to continue . . .`

Kallamamran

I think one thing missing is a progress bar for the generation. Not knowing if it has done 1% or 90% is quite crucial :)

SECourses

the other post is public for people to see. this is only because of Patreon. they dont allow us to make good post with proper setup like zip file only downloadable to paid members. I complain to them all to time but they keep ignoring me

SECourses

I just made update to V8. can you update and try again? it should fix. let me know pls

SECourses

hi you didnt download models :D please donwload models

hacbachvotinh

Can you set up a custom mode that allows pausing intermittently during video processing? Each time I run the process, the GPU overheats, so I want to introduce breaks during processing. The processing time may increase, but this would help reduce GPU temperature.

SECourses

select your cmd and press CTRL+S it will pause. then select back and hit CTRL+Q it will resume

BecauseReasons

Would be awesome to have Kijai's models for this instead or even better Quantized variants :p

The90sPope

I'm wondering what quality people are getting on 1.3B model? I set steps to 100 and quality to 10, yet every video looks terrible. Looks a lot worse than Hunyuan, not even comparable. Everything is over-sharpened and grainy. Resolution looks lover than 480p as well. In other words - videos are unusable. Interpolation or upscaling would not work here. It's just a mess, even on very simple prompts. Motion, however, is slightly more coherent than Hunyuan, but that does not compensate for videos looking unusable.

cool1

You could also adjust the GPU fan speed settings in the Bios.

SECourses

i think this is faster and higher quality than them :D but i plan to cover swarmui as well so we can compare there

SECourses

i really got good quality results. but it may depend on prompt. did you try 14b model? it is way more powerful. perhaps try and see how it works. i made a tutorial video and shown some examples where. editing video to publish

The90sPope

I have RTX 4080 Super with 16GB VRAM and 32GB system RAM. I can't use 14B model. It will take 14 hours to generate 5 second video on 36 steps at 720p. I am stuck with 1.3B model, and results are not quite on par with Hunyuan. With Hunyuan I can do 960x544 40 steps 20fps 5 second videos in a little over 30min and they look OK, something you can work with.

hacbachvotinh

My system automatically sets the fan speed to 100% when the temperature reaches 82°C or higher. However, even at full fan speed, running intensive tasks for an extended period (e.g., rendering a video that takes 5 minutes) can still push the temperature up to 92-94°C. However, if I break the process into smaller steps, such as rendering 5% of the video and then pausing for 3 seconds to cool down before continuing, the temperature remains around 88°C or lower. This allows me to render multiple videos without worrying about damaging the GPU.

Viktor Nagy

thank you for your work, you must put a lot of time and energy into it:)

The90sPope

I think I know what's wrong, managed to generate good quality video now. Tiled VAE decode option unchecked will cause terrible quality if it crashes: [CMD] Generation complete. Overall Duration: 2217.82 seconds. Last used seed: 480351759 ERROR:asyncio:Exception in callback _ProactorBasePipeTransport._call_connection_lost(None) handle: Traceback (most recent call last): File "C:\Python310\lib\asyncio\events.py", line 80, in _run self._context.run(self._callback, *self._args) File "C:\Python310\lib\asyncio\proactor_events.py", line 165, in _call_connection_lost self._sock.shutdown(socket.SHUT_RDWR) ConnectionResetError: [WinError 10054] An existing connection was forcibly closed by the remote host

Chris Fis

3.5 hours for 81 frame 16fps 720p video with 3090?

SECourses

thank you so much. i have been non stop working for 2 days now :D

SECourses

yes if you uncheck tiled it needs a lot of VRAM

SECourses

yes makes total sense but sadly pipeline doesn't support it. you can undervolt instead

The90sPope

Is it possible to use it with dual GPU on the same PC? I have 4080super, and I'm also considering grabbing a used 3090. Would that even work or would I be still limited to 4080?

Pak26

I tried redownload only the i2v 480 model but the download crash because I was getting out of hdd space. So I decided to delete my entire wan_gen file to recover all my previous hdd space, but even after that I still have lost almost 30gb to 40gb of hdd space. Do you have an idea where I could find on my C drive those leftover files ? Thanks

macmotu

I heard about a backdoor found in the WAN AI repo. Is this a problem for me, as I‘ve already used your installer?

weirdgoat

I can't get I2V to work with 1.3b, it always resorts to t2v. Am i missing something? Or ist it t2v/v2v only? 14b i2v 480p makes windows go "end my suffering" and ends in a mem-caused bluescreen, or crashed the App without an error or comment. Running on RTX 4080 with 32G sys mem

EvanGuo

WARNING: Failed to find CUDA.

cool1

So it's sending telemetry to hugging face and/or gradio (or telemetry in visual studio)? Can't those be disabled or have they been? Is that why it sometimes gives connection errors if the router is turned off on some gradio apps part way through processing - because it's trying to send telemetry somewhere? What info does it send? There's an analytics.py in the gradio folder(s) of things like Wan 2.1 video generator. Can't stuff like that be disabled in each app? in .bat files that create a virtual environment and then call the main app, can we just add a line about settings analytics to disabled, like edit Start_App.py of wan 2.1 to: @echo off cd Wan2.1 call .\venv\Scripts\activate.bat || exit /b set GRADIO_ANALYTICS_ENABLED=False // <- this new bit? etc....

cool1

For a simple scene (without complex human motion) it seems easier and quicker to use Cosmos ai video as that will generate 720p video of about 121 frames in about 23 mins on 3090, a lot faster (that I used with image to video). Though maybe it won't be as accurate for some things. Though the 1.3B parameter low res text to video model of Wan 2.1 was a lot faster at maybe 7 or 8 mins or maybe more depending on nr of frames. Though I didn't get great quality output from most of the tests of that (eg. most when I asked it for line drawing style animation). Though it was better when asked for a painting or more real style.

SECourses

1.3b is only for text to video and video to video sadly. they didnt make it for image to video :( . to use 14b models you need to upgrade ram. before upgrading try to load as fp8 and increase your virtual ram. it may work

SECourses

yes you can set environment and disable it true. i really dont bother with it

SECourses

yes each model have different strong aspects

weirdgoat

fp8 causes Bluescreens. I ended up renting some H100 machine, works good, however using fp8 on h100 seems to cause color loss, green layering and weird artifacts

puk

Make a new Install CUDA. This resolve my Problem

Paul Gardiner

Hi, I've the same issue : Geforce rtx 360ti

Paul Gardiner

HI I've installed it twice, first time interface wouldnt open, It will now load the interface but not render anything, I've completed runpod_install, Windows install and windows download models, download ubuntu, Thanks

Paul Gardiner

WARNING: Failed to find CUDA. * Running on local URL: http://127.0.0.1:7860 To create a public link, set `share=True` in `launch()`. [CMD] Loading model: 1.3B with torch dtype: torch.bfloat16 and num_persistent_param_in_dit: 12000000000 Loading models from: models/Wan-AI/Wan2.1-T2V-1.3B/diffusion_pytorch_model.safetensors model_name: wan_video_dit model_class: WanModel This model is initialized with extra kwargs: {'model_type': 't2v', 'patch_size': (1, 2, 2), 'text_len': 512, 'in_dim': 16, 'dim': 1536, 'ffn_dim': 8960, 'freq_dim': 256, 'text_dim': 4096, 'out_dim': 16, 'num_heads': 12, 'num_layers': 30, 'window_size': (-1, -1), 'qk_norm': True, 'cross_attn_norm': True, 'eps': 1e-06}

SECourses

you didnt answer my question. is this on windows? ubuntu? where?

SECourses

well on h100 absolutely no reason to use fp8 :D

SECourses

hi support arrived. can you make a fresh install and let me know? you can copy paste models

SECourses

hi support arrived. can you make a fresh install and let me know? you can copy paste models

Nabil Boulezaz

How many times did u update this i dont get it i just download it like 3 days ago

SECourses

if working atm you dont need to update :) you can just run Windows_Update.bat file. but i keep updating as something new happens . but time to time fresh install can be good because depended libraries are getting updated as well - you can always move models

Paul Gardiner

Sorry, didn't pick up on the question, it's windows, I thought you meant did I click on the Ubuntu file.

SECourses

ok please list me which requirements you installed. python version cuda cudnn c++ tools?

EvanGuo

I have two cards in my PC. one is rtx8000,the other one is rtx4090

EvanGuo

Checkpoint files will always be loaded safely. Total VRAM 24564 MB, total RAM 98066 MB pytorch version: 2.6.0+cu124 Set vram state to: NORMAL_VRAM Device: cuda:0 NVIDIA GeForce RTX 4090 : cudaMallocAsync WARNING: Failed to find CUDA. Using pytorch attention ComfyUI version: 0.3.18 [Prompt Server] web root: E:\meger\ComfyUI\web [ComfyUI-Easy-Use] server: v1.2.7 Loaded

EvanGuo

comfyui cannot find

Sirens Talent Agency

when I'm generating an i2v 14b my task managers shows that my ram is being used and not my vram. then it says connection errored out and the cmd window closes. Suggestions?

Sugar Coat VFX Design

Thank you for this and the Youtube tutorial video. Will you add Lora Support later?

SECourses

yes hopefully i will add. i plan to make training tutorial too

SECourses

how much RAM you have? it needs a lot for 14B models. increase your virtual RAM and try again but you need a lot of RAM. also load as FP8

cool1

When I'm running batch mode image 2 to video with 24 GB preset, 14B image to video 480p, torch.float16, 50 inf steps, 81 frames, 2 of approx 13 output videos (I did start the batch run again early on so some might not be from that batch) were completely black videos (about 15%). This is despite the source images and their .txt files seeming to be all very similar to each other (both of the vids that have all black have valid .txt files with caption and valid source image). The cmd window doesn't show any error for any of the black ones. Though the gradio window shows "Error" on the batch status log window (I don't know if that was because of it trying to connect or if it's some other error). Maybe it could just be a low random probability that it happens with this model.

SECourses

probably model related errors i assume. because it uses exact same mechanism . also very fresh install could fix it because library got some updates and bug fixes

SECourses

hi this gradio doesnt use ComfyUI but from your message i see comfyui doesnt support rtx8000

VDR

Need a LORA :) Hope, you'll do it )

Paul Gardiner

I installed all the apps in your video for trying out previous softwares you provided, Visomaster being the last one, the Wan 2.1 app interface loads but it wont render anything the timer runs, I ran it for over an hour last time but nothing was produced, it says error when I click on it.

cool1

One place talked about lora training for it in comfy, though it mentioned training individual image files with .txt captions (like for an image model). Maybe that would work well for what someone looks like but won't that not be able to capture the motion of what the lora is trained on like training it on video would? Or is it possible for both (image and video training for it)? And would training work locally with 24 GB VRAM?

cool1

Do you know more about the update that's supposed to significantly speed it up - especially the 720p one (the attention one). Are they still working on that?

SECourses

yes hopefully coming soon. please message me a good lora for testing

SECourses

yes i am following. they said they gonna add sageattention too. currently our speed close to comfyui with better quality with full bf16

The90sPope

Updated fresh to v12, but still says v11 and no scaler option.

Draxel

Windows. I'm getting this on V12. File "C:\Windows\System32\Wan2.1\venv\lib\site-packages\triton\runtime\build.py", line 57, in _build raise RuntimeError("Failed to find C compiler. Please specify via CC environment variable.") RuntimeError: Failed to find C compiler. Please specify via CC environment variable.

Simon

Thanks for providing this nice app! I noticed one small bug: When selecting the aspect ratio 9:16_low (e.g. with the 720P model) width and height are wrong. It is preconfigured as 832x480 (same as 16:9_low - not swapped). Previously I used CogVideoX Image2Video with ComfyUI. There it was possible to define a start frame as well as an end frame. Is this possible with WAN, too?

hacbachvotinh

hi, why i can't see Practical-RIFE on V12?

Haroon Aslam

cant see CFG and new panel. something wrong with the update ?

SECourses

Update wasn't pushed. My bad. Just pushed. please update

SECourses

Update wasn't pushed. My bad. Just pushed. please update

SECourses

GGUF speed almost same as ours at the moment we have tested. In some cases GGUF slightly faster we are trying to fix the gap as well

SECourses

Thanks hopefully will fix this. Currently we can't set it as far as I know but I am following the developments

SECourses

I am glad you solved. This also exists in your requirements tutorial source post.

SECourses

Update wasn't pushed. My bad. Just pushed. please update

SECourses

Well I am gonna use same library as the one we use for training - my initial plan. They fully optimize so we will see. Didn't have time yet. First planning to add Lora support

SECourses

I need more info. What do you see in cmd? Installation full logs also needed to be looked.

SECourses

Also good news. Teacache (I made a request for this) may get added and there is a pull request for sage attention. Both can speed up significantly

Pablo Montero

In the installation, 2 errors: ERROR: pip's dependency resolver does not currently take into account all the packages that are installed. This behaviour is the source of the following dependency conflicts. diffsynth 1.1.1 requires transformers==4.46.2, but you have transformers 4.49.0 which is incompatible. ERROR: pip's dependency resolver does not currently take into account all the packages that are installed. This behaviour is the source of the following dependency conflicts. scikit-image 0.25.2 requires numpy>=1.24, but you have numpy 1.23.5 which is incompatible.

SECourses

No issues it works you can ignore

weirdgoat

after the update, prompt enhance works only once, after changing the prompt and running it again the call runs into an error, are you checking if the model is already loaded?

JOHN FIORE

Hello, I got an error trying to download the models, not sure what's going wrong. I followed the requirements correctly and also deleted that one line that makes it download faster, as you advised in your video, but I get the same error: https://pastebin.com/kjGM6sZ4

Juan Grande

Wan21_v12 = ImportError: DLL load failed while importing flash_attn_2_cuda: No se puede encontrar el módulo especificado.

Hockey

I got 64 GB 3,200 MHz RAM, and 16 GB RTX 3080 notebook version. Settings: Model: WAN 2.1 14B Image-to-Video 720P GPU VRAM Preset: 16 GB Aspect Ration: 16:9 Width: 1,280 Height: 780 Tiled VAE Decode: Activated Inference Steps: 50 FPS Multiplier: 4x GPU Settings: torch.bfloat16 Quality: 10 FPS (for saving video): 8 Number of Frames: 81 If not mentioned it means the setting is default. I want the best quality but I'm getting 4811.58/it for 2/50 steps. (2:41:07) Estimated time to finish: ~64 hours. Either something is wrong or my settings are way to high. I don't mind waiting 8 hours for a 5 second video. Any advice if something might be wrong or best quality settings for 8 hours work yielding 5 second video. I made a clean install of Wan21_v12.zip and downloaded models: Wan2.1-I2V-14B-720P Qwen2.5-14B-Instruct

Paul Gardiner

Hi, Reinstalled , followed tutorials for all teh latest requirements but still not working beyond loading the interface. I'd appreciate your advice. To create a public link, set `share=True` in `launch()`. WAN 2.1 1.3B (Text/Video-to-Video) WAN 2.1 1.3B (Text/Video-to-Video) [CMD] Loading model: 1.3B with torch dtype: torch.bfloat16 and num_persistent_param_in_dit: 500000000 Loading models from: models/Wan-AI/Wan2.1-T2V-1.3B/diffusion_pytorch_model.safetensors model_name: wan_video_dit model_class: WanModel This model is initialized with extra kwargs: {'model_type': 't2v', 'patch_size': (1, 2, 2), 'text_len': 512, 'in_dim': 16, 'dim': 1536, 'ffn_dim': 8960, 'freq_dim': 256, 'text_dim': 4096, 'out_dim': 16, 'num_heads': 12, 'num_layers': 30, 'window_size': (-1, -1), 'qk_norm': True, 'cross_attn_norm': True, 'eps': 1e-06} The following models are loaded: ['wan_video_dit']. Loading models from: models/Wan-AI/Wan2.1-T2V-1.3B/models_t5_umt5-xxl-enc-bf16.pth Traceback (most recent call last): File "Z:\Wan21_v12\Wan2.1\venv\lib\site-packages\gradio\queueing.py", line 625, in process_events response = await route_utils.call_process_api( File "Z:\Wan21_v12\Wan2.1\venv\lib\site-packages\gradio\route_utils.py", line 322, in call_process_api output = await app.get_blocks().process_api( File "Z:\Wan21_v12\Wan2.1\venv\lib\site-packages\gradio\blocks.py", line 2108, in process_api result = await self.call_function( File "Z:\Wan21_v12\Wan2.1\venv\lib\site-packages\gradio\blocks.py", line 1655, in call_function prediction = await anyio.to_thread.run_sync( # type: ignore File "Z:\Wan21_v12\Wan2.1\venv\lib\site-packages\anyio\to_thread.py", line 56, in run_sync return await get_async_backend().run_sync_in_worker_thread( File "Z:\Wan21_v12\Wan2.1\venv\lib\site-packages\anyio\_backends\_asyncio.py", line 2461, in run_sync_in_worker_thread return await future File "Z:\Wan21_v12\Wan2.1\venv\lib\site-packages\anyio\_backends\_asyncio.py", line 962, in run result = context.run(func, *args) File "Z:\Wan21_v12\Wan2.1\venv\lib\site-packages\gradio\utils.py", line 890, in wrapper response = f(*args, **kwargs) File "Z:\Wan21_v12\Wan2.1\App.py", line 349, in generate_videos loaded_pipeline = load_wan_pipeline(model_choice, torch_dtype, vram_value) File "Z:\Wan21_v12\Wan2.1\App.py", line 683, in load_wan_pipeline model_manager.load_models( File "z:\wan21_v12\wan2.1\diffsynth-studio\diffsynth\models\model_manager.py", line 422, in load_models self.load_model(file_path, model_names, device=device, torch_dtype=torch_dtype) File "z:\wan21_v12\wan2.1\diffsynth-studio\diffsynth\models\model_manager.py", line 404, in load_model if model_detector.match(file_path, state_dict): File "z:\wan21_v12\wan2.1\diffsynth-studio\diffsynth\models\model_manager.py", line 165, in match if len(state_dict) == 0: TypeError: object of type 'NoneType' has no len()

SECourses

hi please download models that you are going to use. we have model downloader use it.

SECourses

hi try 480p model for now. i am working on speed up hopefully. 14b models are just requiring too much VRAM therefore they are slow. 1.3b model should work blazing fast though.

SECourses

hi this is your internet related. your internet service provider or like antivirus. as a solution i recommend install WARP VPN and enable it and try redownload 1.1.1.1

Paul Gardiner

I downloaded the 1.3 and the 14 but I'll delete them and do it again

SECourses

hi with v14 fixed thanks for reporting

Paul Gardiner

I found my mistake, I copied and pasted the downloaded models from my previous install to save having to download them again to the new install and you evidently cant do that, it works now thank you for your help.

SECourses

thanks a lot for info and you are welcome. i am glad you solved

DanO..

Hi. I'm still getting a blue screen forced restart anytime I try any of the 14B models. RTX 3060 12GB VRAM. 1.3 works fine. I'm hoping quantized models will work better. Is that something you have in mind?

JOHN FIORE

Hello again, thanks for the response. I tried disabling my anti virus and firewall and I'm pretty sure I got a different error, it doesn't mention anything about internet connection this time: https://pastebin.com/2h8Ha8ha Do you think it's still the same issue?

jaehyun AI

Runpod is giving me a lot of errors and won't run.

Andrei Croitor

I am getting an error when downloading any of the 14B Models. Even after removing the "speed up" line of code in the bat file. Does this happen to anyone else?

Al H

Cool - bear in mind RAM offload may be an issue for some users, so a smaller model may help, but I'm sure you're considering that. Thanks for all the hard work :)

Goldo 27

I have a 3090 and everytime I try to generate a image to video (didn't try others yet) it just stops (and I get the "Press any key to continue . . ." in the cmd window) and I could make it work in comfyui, I just wanted something cleaner. : * Running on local URL: http://127.0.0.1:7860 To create a public link, set `share=True` in `launch()`. WAN 2.1 1.3B (Text/Video-to-Video) WAN 2.1 14B Image-to-Video 720P WAN 2.1 14B Image-to-Video 480P Qwen/Qwen2.5-14B-Instruct Loading checkpoint shards: 100%|█████████████████████████████████████████████████████████| 8/8 [01:05<00:00, 8.15s/it] [CMD] Loading model: 14B_image_480p with torch dtype: torch.bfloat16 and num_persistent_param_in_dit: 0 Loading models from: ['models/Wan-AI/Wan2.1-I2V-14B-480P/diffusion_pytorch_model-00001-of-00007.safetensors', 'models/Wan-AI/Wan2.1-I2V-14B-480P/diffusion_pytorch_model-00002-of-00007.safetensors', 'models/Wan-AI/Wan2.1-I2V-14B-480P/diffusion_pytorch_model-00003-of-00007.safetensors', 'models/Wan-AI/Wan2.1-I2V-14B-480P/diffusion_pytorch_model-00004-of-00007.safetensors', 'models/Wan-AI/Wan2.1-I2V-14B-480P/diffusion_pytorch_model-00005-of-00007.safetensors', 'models/Wan-AI/Wan2.1-I2V-14B-480P/diffusion_pytorch_model-00006-of-00007.safetensors', 'models/Wan-AI/Wan2.1-I2V-14B-480P/diffusion_pytorch_model-00007-of-00007.safetensors'] model_name: wan_video_dit model_class: WanModel This model is initialized with extra kwargs: {'model_type': 'i2v', 'patch_size': (1, 2, 2), 'text_len': 512, 'in_dim': 36, 'dim': 5120, 'ffn_dim': 13824, 'freq_dim': 256, 'text_dim': 4096, 'out_dim': 16, 'num_heads': 40, 'num_layers': 40, 'window_size': (-1, -1), 'qk_norm': True, 'cross_attn_norm': True, 'eps': 1e-06} Press any key to continue . . .

SECourses

hi sadly this doesnt tell any error. it could be that your RAM failing. can you try 1.3b model? it should work perfect and let me know please

SECourses

in that case try this. if you are using VPN disable. disable antivirus. try Cloudflare WARP vpn - any of these 3 should fix

SECourses

please tell the errors. but very likely broken pod. if you tell me more info like which model which error i can try and fix if any errors. sorry for late reply

SECourses

hi this is still same error. i will upload models for you as a zip file to download at once. should be ready tomorrow hopefully. so you can manually download and extract

SECourses

hi restart means your system not stable and it is filling entire RAM likely. first verify your PC stability. then increase your virtual RAM. and then load as FP8. how much RAM you have? About quantized thing i am working on that on another interface but currently the library we use doesnt work sadly. but we gonna get hopefully flash attention 3 if works and teacache. i am also working on LoRA right now almost done to support LoRAs

JOHN FIORE

Hello again, I got it to work. I reinstalled everything and it seemed to work this time. However when I try and generate anything it keeps saying "error, connection timed out" over and over. What could be the cause of this?

SECourses

hello. do you see any error on CMD? if not restart PC and try with Google Chrome. let me know

DanO..

Great. Keep at it. Your efforts are well worth the patronage.

Ashly

Thank you so much for this! I am up and running with your v14. I am running the 480 models on a 3090. I have left all the settings on default and to output a 5 second video it is taking roughly 1 hour and 20 min. Does this seem really long? Is there something I can do to optimize? Thanks for your time!

JOHN FIORE

Hello again, managed to fix the connection error. However I'm trying to use the 1.3B video to video option and I keep getting this error now: [CMD] Processing video-to-video with input video: C:\Users\John Fiore\AppData\Local\Temp\gradio\70c76d83032dd401c47d780c3159d4e350715ac045e82e03c67af6929faf4864\vlc-record-2025-03-03-20h11m27s-0hi79sql5xuiqkcz6cu9esource.mp4-_cropped.mp4 Only `num_frames % 4 != 1` is acceptable. We round it up to 77. Traceback (most recent call last): File "C:\Wan21_v12\Wan2.1\venv\lib\site-packages\gradio\queueing.py", line 625, in process_events response = await route_utils.call_process_api( File "C:\Wan21_v12\Wan2.1\venv\lib\site-packages\gradio\route_utils.py", line 322, in call_process_api output = await app.get_blocks().process_api( File "C:\Wan21_v12\Wan2.1\venv\lib\site-packages\gradio\blocks.py", line 2108, in process_api result = await self.call_function( File "C:\Wan21_v12\Wan2.1\venv\lib\site-packages\gradio\blocks.py", line 1655, in call_function prediction = await anyio.to_thread.run_sync( # type: ignore File "C:\Wan21_v12\Wan2.1\venv\lib\site-packages\anyio\to_thread.py", line 56, in run_sync return await get_async_backend().run_sync_in_worker_thread( File "C:\Wan21_v12\Wan2.1\venv\lib\site-packages\anyio\_backends\_asyncio.py", line 2461, in run_sync_in_worker_thread return await future File "C:\Wan21_v12\Wan2.1\venv\lib\site-packages\anyio\_backends\_asyncio.py", line 962, in run result = context.run(func, *args) File "C:\Wan21_v12\Wan2.1\venv\lib\site-packages\gradio\utils.py", line 890, in wrapper response = f(*args, **kwargs) File "C:\Wan21_v12\Wan2.1\App.py", line 424, in generate_videos video_data = loaded_pipeline(input_video=video_obj, denoising_strength=denoising_strength, **common_args) File "C:\Wan21_v12\Wan2.1\venv\lib\site-packages\torch\utils\_contextlib.py", line 116, in decorate_context return func(*args, **kwargs) File "c:\wan21_v12\wan2.1\diffsynth-studio\diffsynth\pipelines\wan_video.py", line 233, in __call__ latents = self.scheduler.add_noise(latents, noise, timestep=self.scheduler.timesteps[0]) File "c:\wan21_v12\wan2.1\diffsynth-studio\diffsynth\schedulers\flow_match.py", line 67, in add_noise sample = (1 - sigma) * original_samples + sigma * noise RuntimeError: The size of tensor a (19) must match the size of tensor b (20) at non-singleton dimension 2 Any idea on what's causing it?

Meet Inventor

WAN 2.1 1.3B Text-to-Video working but WAN 2.1 14B Text-to-Video have following error as follow. Anybody have similar issue? Please share! INFO: Could not find files for the given pattern(s). * Running on local URL: http://127.0.0.1:7860 To create a public link, set `share=True` in `launch()`. WAN 2.1 14B Text-to-Video [CMD] Loading model: 14B_text with torch dtype: torch.bfloat16 and num_persistent_param_in_dit: 1000000000 Loading models from: ['models/Wan-AI/Wan2.1-T2V-14B/diffusion_pytorch_model-00001-of-00006.safetensors', 'models/Wan-AI/Wan2.1-T2V-14B/diffusion_pytorch_model-00002-of-00006.safetensors', 'models/Wan-AI/Wan2.1-T2V-14B/diffusion_pytorch_model-00003-of-00006.safetensors', 'models/Wan-AI/Wan2.1-T2V-14B/diffusion_pytorch_model-00004-of-00006.safetensors', 'models/Wan-AI/Wan2.1-T2V-14B/diffusion_pytorch_model-00005-of-00006.safetensors', 'models/Wan-AI/Wan2.1-T2V-14B/diffusion_pytorch_model-00006-of-00006.safetensors'] model_name: wan_video_dit model_class: WanModel This model is initialized with extra kwargs: {'model_type': 't2v', 'patch_size': (1, 2, 2), 'text_len': 512, 'in_dim': 16, 'dim': 5120, 'ffn_dim': 13824, 'freq_dim': 256, 'text_dim': 4096, 'out_dim': 16, 'num_heads': 40, 'num_layers': 40, 'window_size': (-1, -1), 'qk_norm': True, 'cross_attn_norm': True, 'eps': 1e-06} The following models are loaded: ['wan_video_dit']. Loading models from: models/Wan-AI/Wan2.1-T2V-14B/models_t5_umt5-xxl-enc-bf16.pth Traceback (most recent call last): File "C:\wan_gen\Wan21_v14\Wan2.1\venv\lib\site-packages\gradio\queueing.py", line 625, in process_events response = await route_utils.call_process_api( File "C:\wan_gen\Wan21_v14\Wan2.1\venv\lib\site-packages\gradio\route_utils.py", line 322, in call_process_api output = await app.get_blocks().process_api( File "C:\wan_gen\Wan21_v14\Wan2.1\venv\lib\site-packages\gradio\blocks.py", line 2108, in process_api result = await self.call_function( File "C:\wan_gen\Wan21_v14\Wan2.1\venv\lib\site-packages\gradio\blocks.py", line 1655, in call_function prediction = await anyio.to_thread.run_sync( # type: ignore File "C:\wan_gen\Wan21_v14\Wan2.1\venv\lib\site-packages\anyio\to_thread.py", line 56, in run_sync return await get_async_backend().run_sync_in_worker_thread( File "C:\wan_gen\Wan21_v14\Wan2.1\venv\lib\site-packages\anyio\_backends\_asyncio.py", line 2461, in run_sync_in_worker_thread return await future File "C:\wan_gen\Wan21_v14\Wan2.1\venv\lib\site-packages\anyio\_backends\_asyncio.py", line 962, in run result = context.run(func, *args) File "C:\wan_gen\Wan21_v14\Wan2.1\venv\lib\site-packages\gradio\utils.py", line 890, in wrapper response = f(*args, **kwargs) File "C:\wan_gen\Wan21_v14\Wan2.1\App.py", line 343, in generate_videos loaded_pipeline = load_wan_pipeline(model_choice, torch_dtype, vram_value) File "C:\wan_gen\Wan21_v14\Wan2.1\App.py", line 705, in load_wan_pipeline model_manager.load_models( File "c:\wan_gen\wan21_v14\wan2.1\diffsynth-studio\diffsynth\models\model_manager.py", line 422, in load_models self.load_model(file_path, model_names, device=device, torch_dtype=torch_dtype) File "c:\wan_gen\wan21_v14\wan2.1\diffsynth-studio\diffsynth\models\model_manager.py", line 400, in load_model state_dict = load_state_dict(file_path) File "c:\wan_gen\wan21_v14\wan2.1\diffsynth-studio\diffsynth\models\utils.py", line 69, in load_state_dict return load_state_dict_from_bin(file_path, torch_dtype=torch_dtype) File "c:\wan_gen\wan21_v14\wan2.1\diffsynth-studio\diffsynth\models\utils.py", line 83, in load_state_dict_from_bin state_dict = torch.load(file_path, map_location="cpu", weights_only=True) File "C:\wan_gen\Wan21_v14\Wan2.1\venv\lib\site-packages\torch\serialization.py", line 1376, in load return _legacy_load( File "C:\wan_gen\Wan21_v14\Wan2.1\venv\lib\site-packages\torch\serialization.py", line 1611, in _legacy_load return legacy_load(f) File "C:\wan_gen\Wan21_v14\Wan2.1\venv\lib\site-packages\torch\serialization.py", line 1485, in legacy_load tar.extract("storages", path=tmpdir) File "C:\Python310\lib\tarfile.py", line 2091, in extract tarinfo = self.getmember(member) File "C:\Python310\lib\tarfile.py", line 1813, in getmember raise KeyError("filename %r not found" % name) KeyError: "filename 'storages' not found"

SECourses

hello. looks like models were not downloaded properly. i am uploading all models as zip file to manually download

SECourses

hi yes your frame count not matching : Only `num_frames % 4 != 1` is acceptable. We round it up to 77. can you send me your video so i can apply automatic fix or you have to fix your input video to % 4 = 1. or make a very fresh install and it should fix it. move your models

SECourses

what you can do is either reduce 81 frames to lower to get lower duration or reduce step count to like 20. not much else but don't worry I am working on Flash Attention 3 and Teacache (coming very soon hopefully) and especially teacache should bring like 40% speed up I expect

Meet Inventor

where is the URL of all models you zip?

The90sPope

Can we choose which models to download? Nvidia GPUs are optimized for BF16, it is the fastest for NVidia without loosing much quality, everything else I would go fp32, like text encoder or video encoder. At least on 1.3B model since it's already very lightweight there is no point in losing more precision, I would say.

Ashly

Thanks! 20 steps works much faster and is giving better results on output. It would be amazing if there was a feature that took the last frame from a 5 second video and created another 5 second animation and so on and so forth. Have you experimented with anything like this to create longer videos that are consistent?

The90sPope

there should be an option to run highest possible quality 1.3B model.

Bertrand Samimi

hi bravo for your work unfortuneltly on windiws i have this error during install ERROR: pip's dependency resolver does not currently take into account all the packages that are installed. This behaviour is the source of the following dependency conflicts. numba 0.61.0 requires numpy<2.2,>=1.24, but you have numpy 2.2.3 which is incompatible.

SECourses

please download v21 and run windows updater. currently i know there is numpy related error messages. it works. but i am trying to solve that conflict messages as well. still with conflicts it is working

SECourses

hi currently this is highest quality. because we do inference on bf16. so f32 weights and fp16 weights yields exactly same results i compared.

SECourses

1.3b is still FP32 weights. I am waiting FP16 weights to be published. we are doing inference at BF16

SECourses

hi i couldnt upload yet but i changed to FP16 version atm which takes half space. can you try new download? if still fails i will upload single zip files

SECourses

i think that feature can be added however i dont think would be consistent. but there is a repo i follow which will double generation frame count with consistency. they said they will implement it

Bertrand Samimi

thanks for your help but i have deleted all the folder and restart again during the first windows intall process i have this error

Bertrand Samimi

first ERROR: flash_attn-2.7.4.post1-cp310-cp310-win_amd64.whl is not a supported wheel on this platform. second ERROR: deepspeed-0.16.4-cp310-cp310-win_amd64.whl is not a supported wheel on this platform.

Bertrand Samimi

Third ERROR: Invalid requirement: 'triton==3.2.0-git8f9b005b': Expected end or semicolon (after version specifier) triton==3.2.0-git8f9b005b ~~~~~~~^ four ERROR: Could not install packages due to an OSError: [Errno 2] No such file or directory: 'C:\\Users\\inter\\AppData\\Local\\Temp\\pip-install-fpt0ldeb\\flash-attn_ccae1a0ebe814b7788081cd4952b95bf\\csrc\\composable_kernel\\library\\include\\ck\\library\\tensor_operation_instance\\gpu\\grouped_conv_bwd_weight\\device_grouped_conv_bwd_weight_two_stage_xdl_instance.hpp' what is wrong on my install

The90sPope

Alright, sounds good. Any chance we can get a save settings button, so we don't need to retweak the app on restart? Not a priority tho, just something when and if you have time

SECourses

yes hopefully i will add this. so it can reload your last config

The90sPope

Would be nice to get some light on how to train loras for Wan 2.1 and what GPU can pull it off, if 16gb vram model is good enough? Some sort of Lora training app would be awesome. I understand this is extra work therefore I am willing to pay extra if you pull this off for us. Can't work for free. I feel like my 75 Norwegian kroner a month can be upped a bit haha.

SECourses

yes i plan to make a full guide for wan lora training hopefully soon

SECourses

hi you can try v23 it has some new fixes

Josep Antón Velázquez

Hi! Thanks for your great work! Just tryed to install, I followedall the steps in this video (https://www.youtube.com/watch?v=DrhUHnYfwC0) and reinstaled 2 times but I keep geting this error message, hope you can help me: import error: DLL load failed while importing libtriton: Error en una rutina de inicialización de biblioteca de vínculos dinámicos (DLL). A matching Triton is not available, some optimizations will not be enabled Traceback (most recent call last): File "F:\Installs\Wan 2.1 (Vid Open Source Model)\Wan21_v12\Wan2.1\venv\lib\site-packages\xformers\__init__.py", line 57, in _is_triton_available import triton # noqa File "F:\Installs\Wan 2.1 (Vid Open Source Model)\Wan21_v12\Wan2.1\venv\lib\site-packages\triton\__init__.py", line 20, in from .runtime import ( File "F:\Installs\Wan 2.1 (Vid Open Source Model)\Wan21_v12\Wan2.1\venv\lib\site-packages\triton\runtime\__init__.py", line 1, in from .autotuner import (Autotuner, Config, Heuristics, autotune, heuristics) File "F:\Installs\Wan 2.1 (Vid Open Source Model)\Wan21_v12\Wan2.1\venv\lib\site-packages\triton\runtime\autotuner.py", line 9, in from .jit import KernelInterface File "F:\Installs\Wan 2.1 (Vid Open Source Model)\Wan21_v12\Wan2.1\venv\lib\site-packages\triton\runtime\jit.py", line 12, in from ..runtime.driver import driver File "F:\Installs\Wan 2.1 (Vid Open Source Model)\Wan21_v12\Wan2.1\venv\lib\site-packages\triton\runtime\driver.py", line 1, in from ..backends import backends File "F:\Installs\Wan 2.1 (Vid Open Source Model)\Wan21_v12\Wan2.1\venv\lib\site-packages\triton\backends\__init__.py", line 50, in backends = _discover_backends() File "F:\Installs\Wan 2.1 (Vid Open Source Model)\Wan21_v12\Wan2.1\venv\lib\site-packages\triton\backends\__init__.py", line 43, in _discover_backends compiler = _load_module(name, os.path.join(root, name, 'compiler.py')) File "F:\Installs\Wan 2.1 (Vid Open Source Model)\Wan21_v12\Wan2.1\venv\lib\site-packages\triton\backends\__init__.py", line 12, in _load_module spec.loader.exec_module(module) File "F:\Installs\Wan 2.1 (Vid Open Source Model)\Wan21_v12\Wan2.1\venv\lib\site-packages\triton\backends\amd\compiler.py", line 2, in from triton._C.libtriton import ir, passes, llvm, amd ImportError: DLL load failed while importing libtriton: Error en una rutina de inicialización de biblioteca de vínculos dinámicos (DLL).

SECourses

hi if you have cuda 12.4 and c++ tools as in tutorial then solution is easy. close the program. find .triton cache folder inside C:\Users\your_user_name and delete it. it might be hidden. and restart app and it should get fixed make sure to find .triton and delete

Draxel

man, you're smoking through the versions. Thank you!

Ashly

Hi again! Everything is working great except batch process. I started with V14 but used the Windows_Update.bat tonight, which I assume got me up to date. Whenever I run a batch it completes the first video and then I get the error below, any idea why? Thanks for you continued attention to everything, you are amazing! [CMD] Saved auto processed image: auto_pre_processed_images\-000033-305626887.png [CMD] Saved prompt and parameters: H:\AI_Images\20231208-Cyberpunk-Collection-MetalSkin\_Original Selections-ALL\Picks-for-Image-to-Video\Vertical-OutputVideos\-000033-305626887.txt [CMD] Applying Practical-RIFE with multiplier 2x FPS on video H:\AI_Images\20231208-Cyberpunk-Collection-MetalSkin\_Original Selections-ALL\Picks-for-Image-to-Video\Vertical-OutputVideos\-000033-305626887.mp4 [CMD] Running command: "H:\Wan 2-1- AI Video Gen\Wan2.1\venv\Scripts\python.exe" "Practical-RIFE/inference_video.py" --model="H:\Wan 2-1- AI Video Gen\Wan2.1\Practical-RIFE\train_log" --multi=2 --video="H:\AI_Images\20231208-Cyberpunk-Collection-MetalSkin\_Original Selections-ALL\Picks-for-Image-to-Video\Vertical-OutputVideos\-000033-305626887.mp4" --output="H:\AI_Images\20231208-Cyberpunk-Collection-MetalSkin\_Original Selections-ALL\Picks-for-Image-to-Video\Vertical-OutputVideos\improved_-000033-305626887.mp4" Loaded 3.x/4.x HD model. Traceback (most recent call last): File "H:\Wan 2-1- AI Video Gen\Wan2.1\Practical-RIFE\inference_video.py", line 111, in lastframe = next(videogen) File "H:\Wan 2-1- AI Video Gen\Wan2.1\venv\lib\site-packages\skvideo\io\io.py", line 240, in vreader assert _HAS_FFMPEG, "Cannot find installation of ffmpeg." AssertionError: Cannot find installation of ffmpeg. Traceback (most recent call last): File "H:\Wan 2-1- AI Video Gen\Wan2.1\venv\lib\site-packages\gradio\queueing.py", line 625, in process_events response = await route_utils.call_process_api( File "H:\Wan 2-1- AI Video Gen\Wan2.1\venv\lib\site-packages\gradio\route_utils.py", line 322, in call_process_api output = await app.get_blocks().process_api( File "H:\Wan 2-1- AI Video Gen\Wan2.1\venv\lib\site-packages\gradio\blocks.py", line 2108, in process_api result = await self.call_function( File "H:\Wan 2-1- AI Video Gen\Wan2.1\venv\lib\site-packages\gradio\blocks.py", line 1655, in call_function prediction = await anyio.to_thread.run_sync( # type: ignore File "H:\Wan 2-1- AI Video Gen\Wan2.1\venv\lib\site-packages\anyio\to_thread.py", line 56, in run_sync return await get_async_backend().run_sync_in_worker_thread( File "H:\Wan 2-1- AI Video Gen\Wan2.1\venv\lib\site-packages\anyio\_backends\_asyncio.py", line 2461, in run_sync_in_worker_thread return await future File "H:\Wan 2-1- AI Video Gen\Wan2.1\venv\lib\site-packages\anyio\_backends\_asyncio.py", line 962, in run result = context.run(func, *args) File "H:\Wan 2-1- AI Video Gen\Wan2.1\venv\lib\site-packages\gradio\utils.py", line 890, in wrapper response = f(*args, **kwargs) File "H:\Wan 2-1- AI Video Gen\Wan2.1\App.py", line 750, in batch_process_videos subprocess.run(cmd, shell=True, check=True, env=os.environ) File "C:\Program Files\Python310\lib\subprocess.py", line 526, in run raise CalledProcessError(retcode, process.args, subprocess.CalledProcessError: Command

Ash Murphy

Hi! I have just gone through your tutorial today and installed the I-2-V 14B 720P model. When I went to launch the app though via Start_APP.bat, nothing happened. No error showing. I did the other steps in the other video tutorial. I can see a command screen momentarily open, then it disappears. I then wondered if the Windows_Install_All_GPUs.bat file should be double clicked. I now seem to be getting this: C:\Python310\python.exe: can't open file 'C:\\Wan21_v25\\Wan2.1\\App.py': [Errno 2] No such file or directory Press any key to continue . . . Any ideas?

Lee

Hi. I have an error at the end of installation and similar error when trying to download models that indicates a possible problem with hugging face authorisation. I wonder if I could ask for your advice please? Error details as follows: 1. End on installation... Returning existing local_dir `Wan2.1\Practical-RIFE` as remote repo cannot be accessed in `snapshot_download` (401 Client Error: Unauthorized for url: https://huggingface.co/api/models/MonsterMMORPG/Practical-RIFE/revision/main (Request ID: Root=1-67c81da3-721a6f1c0058e0c834bc1cb4;9d813e52-3874-4fac-8f0d-e0acc64c6393) Invalid credentials in Authorization header). Wan 2.1 installed check out all messages and save them before close to verify any errors or not later Press any key to continue . . . 2. Downloading models... Enter your choice (1-6): 1 Downloading Wan2.1-T2V-1.3B... Traceback (most recent call last): File "C:\Python310\lib\site-packages\huggingface_hub\utils\_http.py", line 409, in hf_raise_for_status response.raise_for_status() File "C:\Python310\lib\site-packages\requests\models.py", line 1024, in raise_for_status raise HTTPError(http_error_msg, response=self) requests.exceptions.HTTPError: 401 Client Error: Unauthorized for url: https://huggingface.co/api/models/Wan-AI/Wan2.1-T2V-1.3B/revision/main The above exception was the direct cause of the following exception: Traceback (most recent call last): File "C:\Python310\lib\site-packages\huggingface_hub\_snapshot_download.py", line 155, in snapshot_download repo_info = api.repo_info(repo_id=repo_id, repo_type=repo_type, revision=revision, token=token) File "C:\Python310\lib\site-packages\huggingface_hub\utils\_validators.py", line 114, in _inner_fn return fn(*args, **kwargs) File "C:\Python310\lib\site-packages\huggingface_hub\hf_api.py", line 2734, in repo_info return method( File "C:\Python310\lib\site-packages\huggingface_hub\utils\_validators.py", line 114, in _inner_fn return fn(*args, **kwargs) File "C:\Python310\lib\site-packages\huggingface_hub\hf_api.py", line 2519, in model_info hf_raise_for_status(r) File "C:\Python310\lib\site-packages\huggingface_hub\utils\_http.py", line 481, in hf_raise_for_status raise _format(HfHubHTTPError, str(e), response) from e huggingface_hub.errors.HfHubHTTPError: 401 Client Error: Unauthorized for url: https://huggingface.co/api/models/Wan-AI/Wan2.1-T2V-1.3B/revision/main (Request ID: Root=1-67c81a8b-6c744d121597e0b048e42ba2;997e9172-bbbd-4cfb-b6f1-c5b675033f8d) Invalid credentials in Authorization header The above exception was the direct cause of the following exception: Traceback (most recent call last): File "H:\AI_Video_Creation\Wan_AI\Download_Models.py", line 71, in download_model(sys.argv[1]) File "H:\AI_Video_Creation\Wan_AI\Download_Models.py", line 12, in download_model snapshot_download( File "C:\Python310\lib\site-packages\huggingface_hub\utils\_validators.py", line 114, in _inner_fn return fn(*args, **kwargs) File "C:\Python310\lib\site-packages\huggingface_hub\_snapshot_download.py", line 235, in snapshot_download raise LocalEntryNotFoundError( huggingface_hub.errors.LocalEntryNotFoundError: An error happened while trying to locate the files on the Hub and we cannot find the appropriate snapshot folder for the specified revision on the local disk. Please check your internet connection and try again. Press any key to continue . . .

Bertrand Samimi

for all user make attention to follow the prerequiste video if you install all correct version this windows installer works perfect thanks for your help and your works

SECourses

assert _HAS_FFMPEG, "Cannot find installation of ffmpeg." AssertionError: Cannot find installation of ffmpeg. hi you are missing ffmpeg. please follow this video and install ffmpeg : https://youtu.be/DrhUHnYfwC0

ForceX

I just did a clean install with v25. Looks like the VENV folder is getting created in the root folder alongside the installer .bat file instead of the Wan2.1 folder. The Start_App.bat is looking for the VENV folder inside the Wan2.1 folder but it’s completely empty except for the models folder. V14 (my previously installed version) installs properly. Maybe something with the V25 installer?

ForceX

Not sure why. The first time in ran the V25 installer the repository was not being copied into the Wan2.1 directory. I remember seeing an error during the install before torch 2.7. I think it said an Error with RDP or RDC. I deleted everything and re-downloaded the V25 package. Everything is now working properly.

Manuel Garrido Peña

Hello, i am following the tutorial, gradio app loads successfully and selects the right model. I enter a prompt and click generate, however nothing happens. And at the terminal I see the following stack trace: ```Traceback (most recent call last): File "D:\Backup\Proyectos\stable-diffusion\sec_courses_patreon_tutorial\Wan21_v25\Wan2.1\venv\lib\site-packages\gradio\queueing.py", line 625, in process_events response = await route_utils.call_process_api( File "D:\Backup\Proyectos\stable-diffusion\sec_courses_patreon_tutorial\Wan21_v25\Wan2.1\venv\lib\site-packages\gradio\route_utils.py", line 322, in call_process_api output = await app.get_blocks().process_api( File "D:\Backup\Proyectos\stable-diffusion\sec_courses_patreon_tutorial\Wan21_v25\Wan2.1\venv\lib\site-packages\gradio\blocks.py", line 2108, in process_api result = await self.call_function( File "D:\Backup\Proyectos\stable-diffusion\sec_courses_patreon_tutorial\Wan21_v25\Wan2.1\venv\lib\site-packages\gradio\blocks.py", line 1655, in call_function prediction = await anyio.to_thread.run_sync( # type: ignore File "D:\Backup\Proyectos\stable-diffusion\sec_courses_patreon_tutorial\Wan21_v25\Wan2.1\venv\lib\site-packages\anyio\to_thread.py", line 56, in run_sync return await get_async_backend().run_sync_in_worker_thread( File "D:\Backup\Proyectos\stable-diffusion\sec_courses_patreon_tutorial\Wan21_v25\Wan2.1\venv\lib\site-packages\anyio\_backends\_asyncio.py", line 2461, in run_sync_in_worker_thread return await future File "D:\Backup\Proyectos\stable-diffusion\sec_courses_patreon_tutorial\Wan21_v25\Wan2.1\venv\lib\site-packages\anyio\_backends\_asyncio.py", line 962, in run result = context.run(func, *args) File "D:\Backup\Proyectos\stable-diffusion\sec_courses_patreon_tutorial\Wan21_v25\Wan2.1\venv\lib\site-packages\gradio\utils.py", line 890, in wrapper response = f(*args, **kwargs) File "D:\Backup\Proyectos\stable-diffusion\sec_courses_patreon_tutorial\Wan21_v25\Wan2.1\App.py", line 392, in generate_videos loaded_pipeline = load_wan_pipeline(model_choice, torch_dtype, vram_value, File "D:\Backup\Proyectos\stable-diffusion\sec_courses_patreon_tutorial\Wan21_v25\Wan2.1\App.py", line 823, in load_wan_pipeline model_manager.load_models( File "d:\backup\proyectos\stable-diffusion\sec_courses_patreon_tutorial\wan21_v25\wan2.1\diffsynth-studio\diffsynth\models\model_manager.py", line 422, in load_models self.load_model(file_path, model_names, device=device, torch_dtype=torch_dtype) File "d:\backup\proyectos\stable-diffusion\sec_courses_patreon_tutorial\wan21_v25\wan2.1\diffsynth-studio\diffsynth\models\model_manager.py", line 404, in load_model if model_detector.match(file_path, state_dict): File "d:\backup\proyectos\stable-diffusion\sec_courses_patreon_tutorial\wan21_v25\wan2.1\diffsynth-studio\diffsynth\models\model_manager.py", line 165, in match if len(state_dict) == 0: TypeError: object of type 'NoneType' has no len() ```

SECourses

looks like you didnt download models :)

Ashly

Thanks! I got ffmpeg working now but either that or the new updated changed some behavior of wan 2.1. It now saves two videos, one of them is called "improved", not sure why this is. My main issue now is that yesterday when I started the program it loaded the model once (which takes 10min or so) and then every generation after that it didn't need to load the model again. Now for every generation it reloads the model again. Any idea why this might be happening? This is what it says before the second generation now [in#1 @ 000002763fb4b9c0] Error opening input: No such file or directory Error opening input file ./temp/audio.mkv. Error opening input files: No such file or directory Audio transfer failed. Interpolated video will have no audio [CMD] Practical-RIFE finished. Improved video saved to: outputs\improved_00022.mp4 [CMD] Generation complete. Overall Duration: 1387.18 seconds. Last used seed: 2880374580 [CMD] No LoRA selected. Using base model. [CMD] Loading model: 14B_image_480p with torch dtype

SECourses

are you generating batch image to video? are you on v25?

Ashly

This is happening when I generate a single video now. I don't think it happens when I do batch (but I'm not sure). I used the "Windows_Update.bat" and it said I was up to date so I assume that means I am on v25

SECourses

yes your internet related error. restart pc and try again. turn off antivirus try again. if all fails try cloudlfare warp VPN and try again until it perfectly git clone and install all

SECourses

you use 480p image to video? which model exactly so i can test and see

Ashly

Yes, I am using 480p 24GB on a 3090.

SECourses

I just analyzed this carefully. it deloads model to be able to VAE decode. sadly no other option atm. did you download FP16 models? they are half size so should be 2x faster to load at least

Ashly

Interesting, so when I had ffmpeg disabled yesterday it wasn't causing it to double load. I just downloaded the models from your script for the 480 version, not any others. Are the FP16 in that script? I'll take a look. While I have your attention I have one more question. I like to iterate multiple version of the same prompt with different seeds but I don't see an option in the setup to, say, output 5 versions in a row. Am I missing it or is this not possible right now? Thanks again!

SECourses

delete older models and download again. we have what you want Number of Generations. search it and increase it

Ashly

gezz, don't know how I missed that, haha. I'll redownload, thanks!

Han Has

Hi, i got this msg after i press generate. something because i have 4GB VRAM. "torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 1.96 GiB. GPU 0 has a total capacity of 4.00 GiB of which 1.29 GiB is free. Of the allocated memory 20.13 MiB is allocated by PyTorch, and 1.96 GiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation" even if i use this command "PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True" or chose this option "torch.float8_e4m3fn" it stuck at this weird msg abt flash attn. RuntimeError: FlashAttention only supports Ampere GPUs or newer. Exception raised from mha_varlen_fwd at C:\a\f\csrc\flash_attn\flash_api.cpp:542 (most recent call first): 00007FFB37392C2400007FFB37392B80 c10.dll!c10::Error::Error [ @ ] 00007FFB3739169A00007FFB37391640 c10.dll!c10::detail::torchCheckFail [ @ ] 00007FF9984A48C500007FF998495C20 flash_attn_2_cuda.cp310-win_amd64.pyd!c10::ivalue::Object::operator= [ @ ] 00007FF9984B7E9900007FF9984AC450 flash_attn_2_cuda.cp310-win_amd64.pyd!PyInit_flash_attn_2_cuda [ @ ] 00007FF9984B5F9F00007FF9984AC450 flash_attn_2_cuda.cp310-win_amd64.pyd!PyInit_flash_attn_2_cuda [ @ ] 00007FF9984B607400007FF9984AC450 flash_attn_2_cuda.cp310-win_amd64.pyd!PyInit_flash_attn_2_cuda [ @ ] 00007FF99849E60800007FF998495C20 flash_attn_2_cuda.cp310-win_amd64.pyd!c10::ivalue::Object::operator= [ @ ]

SECourses

hi what is your GPU model? if it is older version that doesnt support flash attention sadly wont work

cool1

Thanks for the update. It says it now works with Loras (Wan 2.1) but will we be able to train the Loras for it too? I think that was mentioned previously. Or does that not work for windows currently (training images/video for it)?

SECourses

for training i will make a guide hopefully

The90sPope

Great, good things coming. Would be nice to get a speed up and Lora training tutorial. Loving the app so far. Popping videos in 1.3B with 4080super. Takes 5-6min per 5second video at 832x480.

SECourses

Yep i agree. Hopefully lora tutorial coming and also huge speed up for generations

troyau

Would also love to see a decent guide for Wan lora training for local and runpod.

John

With runpod, im getting this when I click generate after a few minutes of it loading. torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 2.26 GiB. GPU 0 has a total capacity of 22.06 GiB of which 2.18 GiB is free. Process 670454 has 19.87 GiB memory in use. Of the allocated memory 18.12 GiB is allocated by PyTorch, and 1.52 GiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation. See documentation for Memory Management (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables) I see the loading bar for a split second and then this error above pops up. Do I need to change my Pods settings? Im using 200gb container disk and 300gb volume disk EDIT- I had a rubbish pod. Ignore me lol. All working now with a more legitimate one.

Wktra

In which directory do we place the LORAs?

SECourses

Put your LoRA models into LoRAs folder inside main Wan2.1 folder

SECourses

awesome. yes sadly a lot of rubbish pods there are :(

SECourses

yes that is actually what i mean :) LoRA training for Wan

Relaxing Music

I am getting the following error on my rtx3060 ti with 16GB VRAM: RuntimeError: getCudnnDataTypeFromScalarType() not supported for Float8_e4m3fn The complete error is: Traceback (most recent call last): File "C:\Wan21_v28\Wan2.1\venv\lib\site-packages\gradio\queueing.py", line 625, in process_events response = await route_utils.call_process_api( File "C:\Wan21_v28\Wan2.1\venv\lib\site-packages\gradio\route_utils.py", line 322, in call_process_api output = await app.get_blocks().process_api( File "C:\Wan21_v28\Wan2.1\venv\lib\site-packages\gradio\blocks.py", line 2103, in process_api result = await self.call_function( File "C:\Wan21_v28\Wan2.1\venv\lib\site-packages\gradio\blocks.py", line 1650, in call_function prediction = await anyio.to_thread.run_sync( # type: ignore File "C:\Wan21_v28\Wan2.1\venv\lib\site-packages\anyio\to_thread.py", line 56, in run_sync return await get_async_backend().run_sync_in_worker_thread( File "C:\Wan21_v28\Wan2.1\venv\lib\site-packages\anyio\_backends\_asyncio.py", line 2461, in run_sync_in_worker_thread return await future File "C:\Wan21_v28\Wan2.1\venv\lib\site-packages\anyio\_backends\_asyncio.py", line 962, in run result = context.run(func, *args) File "C:\Wan21_v28\Wan2.1\venv\lib\site-packages\gradio\utils.py", line 890, in wrapper response = f(*args, **kwargs) File "C:\Wan21_v28\Wan2.1\App.py", line 686, in generate_videos video_data = loaded_pipeline(input_image=processed_image, **common_args) File "C:\Wan21_v28\Wan2.1\venv\lib\site-packages\torch\utils\_contextlib.py", line 116, in decorate_context return func(*args, **kwargs) File "C:\Wan21_v28\Wan2.1\DiffSynth-Studio\diffsynth\pipelines\wan_video.py", line 248, in __call__ image_emb = self.encode_image(input_image, num_frames, height, width) File "C:\Wan21_v28\Wan2.1\DiffSynth-Studio\diffsynth\pipelines\wan_video.py", line 156, in encode_image clip_context = self.image_encoder.encode_image([image]) File "C:\Wan21_v28\Wan2.1\DiffSynth-Studio\diffsynth\models\wan_video_image_encoder.py", line 879, in encode_image out = self.model.visual(videos, use_31_block=True) File "C:\Wan21_v28\Wan2.1\venv\lib\site-packages\torch\nn\modules\module.py", line 1751, in _wrapped_call_impl return self._call_impl(*args, **kwargs) File "C:\Wan21_v28\Wan2.1\venv\lib\site-packages\torch\nn\modules\module.py", line 1762, in _call_impl return forward_call(*args, **kwargs) File "C:\Wan21_v28\Wan2.1\DiffSynth-Studio\diffsynth\models\wan_video_image_encoder.py", line 460, in forward x = self.patch_embedding(x).flatten(2).permute(0, 2, 1) File "C:\Wan21_v28\Wan2.1\venv\lib\site-packages\torch\nn\modules\module.py", line 1751, in _wrapped_call_impl return self._call_impl(*args, **kwargs) File "C:\Wan21_v28\Wan2.1\venv\lib\site-packages\torch\nn\modules\module.py", line 1762, in _call_impl return forward_call(*args, **kwargs) File "C:\Wan21_v28\Wan2.1\DiffSynth-Studio\diffsynth\vram_management\layers.py", line 38, in forward return module(*args, **kwargs) File "C:\Wan21_v28\Wan2.1\venv\lib\site-packages\torch\nn\modules\module.py", line 1751, in _wrapped_call_impl return self._call_impl(*args, **kwargs) File "C:\Wan21_v28\Wan2.1\venv\lib\site-packages\torch\nn\modules\module.py", line 1762, in _call_impl return forward_call(*args, **kwargs) File "C:\Wan21_v28\Wan2.1\venv\lib\site-packages\torch\nn\modules\conv.py", line 554, in forward return self._conv_forward(input, self.weight, self.bias) File "C:\Wan21_v28\Wan2.1\venv\lib\site-packages\torch\nn\modules\conv.py", line 549, in _conv_forward return F.conv2d( RuntimeError: getCudnnDataTypeFromScalarType() not supported for Float8_e4m3fn And seems to work when I select "torch.bfloat16"

SECourses

hi. make a fresh install it will fix. make fresh install and move models. i just tested on rtx 3090 ti which is basically same genre card

Bjoern

Thanks for this. I am curious, why there is a Video-2-Video 1.3b Version for low VRAM, but no Image-2-Video for 1.3b? Shouldn't image-2-video be "easier" to handle and less ressources consuming than video-2-video?

SECourses

excellent question. sadly wan ai team didn't publish such model yet :(

Bjoern

Thanks for the quick reply. Hope they will release something in the near future. But I am sure you will keep us updated.

Morgan Old

Thank you very much, as always, very useful and practical. My video card 4090 generates 720 5 second video in 30 steps for about 50 minutes, I don't know if this is normal or not? Today I tried v2v and got this error ***** Only `num_frames % 4 != 1` is acceptable. We round it up to 173. Traceback (most recent call last): File "C:\fr_work\wan_2_1\Wan2.1\venv\lib\site-packages\gradio\queueing.py", line 625, in process_events response = await route_utils.call_process_api( File "C:\fr_work\wan_2_1\Wan2.1\venv\lib\site-packages\gradio\route_utils.py", line 322, in call_process_api output = await app.get_blocks().process_api( File "C:\fr_work\wan_2_1\Wan2.1\venv\lib\site-packages\gradio\blocks.py", line 2103, in process_api result = await self.call_function( File "C:\fr_work\wan_2_1\Wan2.1\venv\lib\site-packages\gradio\blocks.py", line 1650, in call_function prediction = await anyio.to_thread.run_sync( # type: ignore File "C:\fr_work\wan_2_1\Wan2.1\venv\lib\site-packages\anyio\to_thread.py", line 56, in run_sync return await get_async_backend().run_sync_in_worker_thread( File "C:\fr_work\wan_2_1\Wan2.1\venv\lib\site-packages\anyio\_backends\_asyncio.py", line 2461, in run_sync_in_worker_thread return await future File "C:\fr_work\wan_2_1\Wan2.1\venv\lib\site-packages\anyio\_backends\_asyncio.py", line 962, in run result = context.run(func, *args) File "C:\fr_work\wan_2_1\Wan2.1\venv\lib\site-packages\gradio\utils.py", line 890, in wrapper response = f(*args, **kwargs) File "C:\fr_work\wan_2_1\Wan2.1\App.py", line 657, in generate_videos video_data = loaded_pipeline(input_video=video_obj, denoising_strength=denoising_strength, **common_args) File "C:\fr_work\wan_2_1\Wan2.1\venv\lib\site-packages\torch\utils\_contextlib.py", line 116, in decorate_context return func(*args, **kwargs) File "C:\fr_work\wan_2_1\Wan2.1\DiffSynth-Studio\diffsynth\pipelines\wan_video.py", line 235, in __call__ latents = self.scheduler.add_noise(latents, noise, timestep=self.scheduler.timesteps[0]) File "C:\fr_work\wan_2_1\Wan2.1\DiffSynth-Studio\diffsynth\schedulers\flow_match.py", line 67, in add_noise sample = (1 - sigma) * original_samples + sigma * noise RuntimeError: The size of tensor a (43) must match the size of tensor b (44) at non-singleton dimension 2 *****

SECourses

50 min normal but v2v shouldnt given error. can you send me video as a zip so i can fix? these errors happens due to your video doesnt have accurate number of frames or not matching your target frame. i will try to make it auto fixed : monstermmorpg@gmail.com

SECourses

sure i am following. people asked this

The90sPope

Where can we get WAN Loras? A LoRa training app or addition to existing app would be nice. Also, GGUF versions for 14B would be nice. I am able to use 14B T2V Q6 on 16gb vram in comfyui and it takes around 22 minutes for 480p 5s video. But 14B 480p and 1.3B 480p have big difference. On 14B 480p looks much nicer, faces are clear, basically everything is drastically better.

SECourses

GGUF sacrifices quality. it is a tradeoff

Chris B

Excellent work! Are there any recommendations for the best number of inference steps? Right now, on a 4090, 50 steps take about ~25 minutes using default settings for 480p image to video. Also, if you increase frames, that will increase duration. How long of a duration has been tested to see how longer videos turn out? Do you need more inference + frames to get a good result? This is one thing I haven't tested yet. Looking forward to speed improvements hopefully coming soon!

SECourses

thanks. best one is 81 frames. lower can be made. a little bit higher also. 50 steps great but work as low as 20 too (reduce quality sadly)

Chris B

What's the highest frames you've tested?

SECourses

81 so far which is native count :D

GREG TAYLOR

Every time I try to generate a image I get a Error message that say connection timed out after a couple seconds of generating. Can you tell me how to fix this. The first time I tried to generate a image it just kept working on it but never seen a image generated I closed the Wan 2.1 and restarted and that is when I kept getting errors. Try different setting but no change.

SECourses

well it is related to your windows setup - happens with gradio. sadly i dont know exact solution but restarting pc and starting only 1 gradio may help

The90sPope

At 16gb vram I'm basically stuck with 1.3B, which has worse quality compared to 14B GGUF Q6 which I'm able to run. It pops a 480p 5s video in 20mins, while regular 14B would overflow to RAM and hardrive making it multiple hours for 5s video. So GGUF is a solution for 16gb vram. I can get a lot better quality with 14B GGUF compared to 1.3B model and the inference time is OKish. 1.3B spits a video out in 5 to 6 minutes, GGUF does it in 20min but the extra time is worth it. I wish I could run 14B normal model, it's amazing.

SECourses

How much ram you have? Did you try fp8? 16 gb can run 13b 480p model fairly well

The90sPope

32gb system RAM. It does run 1.3B no problem, but 14B takes many hours for 5s video, while 14B GGUF Q6 takes only 20min and the quality is way better than 1.3B

Fco Muñoz

Hi in Massed Compute i install and download the models with no problem but when i want to start the app it gives me an error source venv/bin/activate bash: venv/bin/activate: No such file or directory i checked it and indeed that file is not there.

SECourses

reinstall. if you make a video of how you are doing i can pinpoint your error

SECourses

yep 32 GB is the biggest problem. i will test if our app can run GGUF lets see

Bradley Baker

Same issue on a fresh install on windows

Josh spivey

Traceback (most recent call last): File "Download_Models.py", line 4, in from huggingface_hub import snapshot_download ModuleNotFoundError: No module named 'huggingface_hub'

Fco Muñoz

[CMD] No LoRA selected. Using base model. [CMD] Loading model: 14B_image_720p with torch dtype: torch.bfloat16 and num_persistent_param_in_dit: 0 Loading models from: ['models/Wan-AI/Wan2.1-I2V-14B-720P/diffusion_pytorch_model-00001-of-00007.safetensors', 'models/Wan-AI/Wan2.1-I2V-14B-720P/diffusion_pytorch_model-00002-of-00007.safetensors', 'models/Wan-AI/Wan2.1-I2V-14B-720P/diffusion_pytorch_model-00003-of-00007.safetensors', 'models/Wan-AI/Wan2.1-I2V-14B-720P/diffusion_pytorch_model-00004-of-00007.safetensors', 'models/Wan-AI/Wan2.1-I2V-14B-720P/diffusion_pytorch_model-00005-of-00007.safetensors', 'models/Wan-AI/Wan2.1-I2V-14B-720P/diffusion_pytorch_model-00006-of-00007.safetensors', 'models/Wan-AI/Wan2.1-I2V-14B-720P/diffusion_pytorch_model-00007-of-00007.safetensors'] model_name: wan_video_dit model_class: WanModel This model is initialized with extra kwargs: {'has_image_input': True, 'patch_size': [1, 2, 2], 'in_dim': 36, 'dim': 5120, 'ffn_dim': 13824, 'freq_dim': 256, 'text_dim': 4096, 'out_dim': 16, 'num_heads': 40, 'num_layers': 40, 'eps': 1e-06} Killed image 2 video

SECourses

hello out of RAM. please set 100 GB virtual RAM on your fastest disk and try again and let me know.

SECourses

hello. open a cmd and execute below commands huggingface-cli logout && unset HUGGINGFACE_TOKEN HUGGINGFACE_HUB_TOKEN also pip install huggingface_hub hf_transfer --upgrade

Ammar Salem

When I run Windows_Download_Models.bat I am getting error: Enter your choice (e.g., 1,3,4,5): 6 [INFO] Downloading selected model(s)... Phase 1: Downloading specific files for Wan-AI/Wan2.1-T2V-1.3B from Wan-AI/Wan2.1-T2V-1.3B ... Traceback (most recent call last): File "G:\AI_Projects_G\Wan2.1\Download_Models.py", line 171, in main() File "G:\AI_Projects_G\Wan2.1\Download_Models.py", line 159, in main download_model_in_two_phases(repo_id, model_folder, common_items) File "G:\AI_Projects_G\Wan2.1\Download_Models.py", line 22, in download_model_in_two_phases snapshot_download(repo_id=repo_id, local_dir=model_specific_dir, ignore_patterns=common_items) File "C:\Python310\lib\site-packages\huggingface_hub\utils\_validators.py", line 120, in _inner_fn return fn(*args, **kwargs) File "C:\Python310\lib\site-packages\huggingface_hub\_snapshot_download.py", line 233, in snapshot_download _inner_hf_hub_download(file) File "C:\Python310\lib\site-packages\huggingface_hub\_snapshot_download.py", line 212, in _inner_hf_hub_download return hf_hub_download( File "C:\Python310\lib\site-packages\huggingface_hub\utils\_validators.py", line 120, in _inner_fn return fn(*args, **kwargs) File "C:\Python310\lib\site-packages\huggingface_hub\file_download.py", line 1280, in hf_hub_download return _to_local_dir(blob_path, local_dir, relative_filename, use_symlinks=local_dir_use_symlinks) File "C:\Python310\lib\site-packages\huggingface_hub\file_download.py", line 1568, in _to_local_dir _create_symlink(real_blob_path, local_dir_filepath, new_blob=False) File "C:\Python310\lib\site-packages\huggingface_hub\file_download.py", line 873, in _create_symlink _support_symlinks = are_symlinks_supported(os.path.dirname(os.path.commonpath([abs_src, abs_dst]))) File "C:\Python310\lib\ntpath.py", line 807, in commonpath raise ValueError("Paths don't have the same drive") ValueError: Paths don't have the same drive Download process completed. Press any key to exit.

cool1

There's a "prompt enhance" option in this where it tries to enhance whatever prompt you enter, but it doesn't see what's in the image I think. And there's no control over how much extra text it adds. There's another app and model you gave (CogVLM2 image captioner) that is multi-modal and can see the image you give, so you could ask it to create a video prompt from that (custom instruction instead of just to describe the image) of what it thinks could happen in a short 5 second-ish video where the given image is the 1st frame of the video. You could also specify how many tokens to use (which you can't for the current one). So it would be good if there could also be an option for that, that used that multi-modal AI model to look at the image and create the video prompt from it.

SECourses

hi install Wan 2.1 into C drive and it will work. later you can reinstall into G drive and move models

Ammar Salem

Hey, thanks for the pointer. I added in system env under user env Variable name: HF_HOME Variable value: G:\huggingface_cache and it solved it ;-) Thanks a lot man!

SECourses

yes i plan to add all loras to downloader. they are not big so all can be downloaded :D

Ritesh 'Tez' B

i think there is an issue with the export when i use I2V 480p. no matter what aspect ratio is picked even if the pre processor image is the right resolution and aspect ratio both the outputs are forcing a 16:9. having this issue with only the latest v32. my tests only was with 3:4 ratio dont know if it is the same issue on others didnt get to check it yet. On a side note can you look into diffusion-Pipe wan 2.1 training lora i found a GUI on civitai but its in chinese. The link is below https://civitai.com/articles/10335/hunyuan-video-lora-trainning-with-gui-in-windows

cool1

When I downloaded the latest version 33 .bat file and extract all the files (except I keep the previous start_app.bat as that hasn't changed, so I keep the gradio analytics off as the newest versions never have it turned off, which will constantly send data to somewhere online). I clicked "update windows.bat" and that seemed to work okay. Then when I run the app and click generate with a 14b image to video, with rife unchecked (as that said it couldn't find ffmpeg, but I can interpolate in another app if needed), with the new teacache enabled, it says: "(in Wan2.1\venv\lib\site-packages\torch\utils\_contextlib.py) line 116, in decorate_context return func(*args, **kwargs) TypeError: WanVideoPipeline.__call__() got an unexpected keyword argument 'tea_cache_l1_thresh'" So now it doesn't create a video. But with the version (about version 31) with your teacache it did work.

Ammar Salem

Great! And please consider adding predefined prompts for each lora to load on selection 🙏

SECourses

make a fresh error obviously update failed for some reason and move models

cool1

a fresh installation and cutting the models out and then pasting them back? I didn't see any errors when "update windows" was run though but I'll try that thanks.

SECourses

yep fresh install and cut back models

SECourses

sorry error fixed with v33. and i plan to make a better training workflow hopefully than that

SECourses

yes certainly i can add trigger words will do :)

cool1

I've done that. I tried using it but it didnt' seem as fast as before but maybe it was things like the teacache setting. In Teacache L1 treshold if I set that to 0.15 (because it says "0.05 Threshold for 1.3b model and 0.15 for 14b models recommended") and then press "save config" it saves ""teacache_thresh": 0.15" in the current config file. Then if I press "load config" even though it's set at 0.15 in the current config file it sets teacache_thresh back to 0.05. Also the Tea Cache Model ID is set to "Wan2.1-T2V-1.3B" but the "model and resolution" is set to WAN 2.1 14B Image-to-Video 480P. I don't know why it's using a different model there in teacache model id than it is in the model and resolution" setting. It worked a lot faster when I re-ran it (with the 0.15 setting for teacache). But I don't know what that setting isn't getting loaded from the config.

Chad Lumley

Would it be possible to display a dynamic preview image in the Gradio interface? I know you can do it in ComfyUI and I find it really useful, especially if the video has unexpected movement you can catch early.

Ritesh 'Tez' B

Thank you for the fast fix, and really looking forward to the lora training for it. Not many useful loras on Civitai for wan currently would love to be able to train my own and share with the community.

100Sachen80

Hi Furkan, on my 3060 12GB it takes several hours to generate the video, i think it was 4 or 5 hours. Default settings. 10GB vram preset selected. Teacache enabled. Is it normal to take so long? 480p image to video version.

SECourses

yes 3060 is weak. you need to reduce number of frames and steps.

SECourses

I will research this if possible

SECourses

make a fresh config olders may not work . delete older ones.

cool1

Thanks. I did. I deleted the "wan 2.1" folder (after renaming it to something else so I could copy the models from it. I didn't delete the folder that has the wan21_33.zip file and the "windows_install.bat" file etc. as they're needed to install it and call what gets put in the "wan 2.1" subfolder of that.

Bernice Crample

Same error here after a few installs. Just noticed I didn't RTFM and am using CUDA 11.8 when it clearly says it requires CUDA 12. Intalling 12.8 now. will report back.

SECourses

ye hopefully will cover it fully with easiest way

Bernice Crample

100% yes. Check CUDA version with nvcc --version and update CUDA to 12 like the documentation says to solve: TypeError: WanVideoPipeline.__call__() got an unexpected keyword argument 'tea_cache_l1_thresh'"

SECourses

your installation is outdated. please make a fresh , fresh install and move models folder

George Gostyshev

Great job! Is it possible to add a couple of things into app such as Reuse Seed and some kind of bar to check how much VRAM it currently used\remains?

SECourses

you mean save seed in config? vram usage monitors usually not working great i recommend nvitop

Eugene Pitts

Thank you so much for your work - My text to video works but when I try to do Image to video with the Squish LoRa I get this error: Traceback (most recent call last): File "D:\Wan21_v38\Wan2.1\venv\lib\site-packages\gradio\queueing.py", line 625, in process_events response = await route_utils.call_process_api( File "D:\Wan21_v38\Wan2.1\venv\lib\site-packages\gradio\route_utils.py", line 322, in call_process_api output = await app.get_blocks().process_api( File "D:\Wan21_v38\Wan2.1\venv\lib\site-packages\gradio\blocks.py", line 2103, in process_api result = await self.call_function( File "D:\Wan21_v38\Wan2.1\venv\lib\site-packages\gradio\blocks.py", line 1650, in call_function prediction = await anyio.to_thread.run_sync( # type: ignore File "D:\Wan21_v38\Wan2.1\venv\lib\site-packages\anyio\to_thread.py", line 56, in run_sync return await get_async_backend().run_sync_in_worker_thread( File "D:\Wan21_v38\Wan2.1\venv\lib\site-packages\anyio\_backends\_asyncio.py", line 2461, in run_sync_in_worker_thread return await future File "D:\Wan21_v38\Wan2.1\venv\lib\site-packages\anyio\_backends\_asyncio.py", line 962, in run result = context.run(func, *args) File "D:\Wan21_v38\Wan2.1\venv\lib\site-packages\gradio\utils.py", line 890, in wrapper response = f(*args, **kwargs) File "D:\Wan21_v38\Wan2.1\App.py", line 749, in generate_videos loaded_pipeline = load_wan_pipeline(model_choice, torch_dtype, vram_value, lora_path=effective_loras, lora_alpha=None) File "D:\Wan21_v38\Wan2.1\App.py", line 1267, in load_wan_pipeline model_manager.load_lora(path, lora_alpha=alpha) File "D:\Wan21_v38\Wan2.1\DiffSynth-Studio\diffsynth\models\model_manager.py", line 388, in load_lora lora.load(model, state_dict, lora_prefix, alpha=lora_alpha, model_resource=model_resource) File "D:\Wan21_v38\Wan2.1\DiffSynth-Studio\diffsynth\models\lora.py", line 31, in wrapper result = func(*args, **kwargs) File "D:\Wan21_v38\Wan2.1\DiffSynth-Studio\diffsynth\models\lora.py", line 629, in load param.data = (param_float + lora_weight).to(param.dtype) RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cuda:0 and cpu!

SECourses

True currently LoRA fails with FP8. So use BF16. I will try to fix if i can

SECourses

hello just fixed. please run updater and it should fix. ty for reporting

George Gostyshev

To check same seed but with some different settings. Like any TXT2IMG have. Yeah using nvitop too but sometimes want to see all in one UI.

SECourses

yes we have seed feature already. uncheck random seed and type a seed you want

Joseph Gentsch

How did you end up fixing this? I have the same issue.

cool1

Could a GUI app based on Musubi-tuner (that seems related to Kohya SS) work for Wan 2.1 Lora training, including for Windows. Another trainer app page said (that was related to that but different) said there's (the other app not Musubi-tuner) wouldn't work on Windows. But the Musubi-tuner page says "Windows compatibility confirmed (Linux compatibility confirmed by community)". So maybe it would work, including for Windows PCs (depending on the GPU VRAM and the RAM. It says "VRAM: 12GB or more recommended for image training, 24GB or more for video training" and "Main Memory: 64GB or more recommended, 32GB + swap may work". Would that be workable? Maybe it could also be made more memory efficient for those who have less RAM (though I have enough). See: https://github.com/kohya-ss/musubi-tuner

SECourses

exactly on my list. an amazing gradio app to train with musubi-tuner

cool1

Thanks a lot

Walker4k

Any ideas? - To create a public link, set `share=True` in `launch()`. [CMD] Loading model: 1.3B with torch dtype: torch.bfloat16 and num_persistent_param_in_dit: 12000000000 [WARNING] Neither models\models_t5_umt5-xxl-enc-bf16.pth nor models\Wan-AI\Wan2.1-T2V-1.3B\models_t5_umt5-xxl-enc-bf16.pth found. Using models\Wan-AI\Wan2.1-T2V-1.3B\models_t5_umt5-xxl-enc-bf16.pth as fallback. [WARNING] Neither models\Wan2.1_VAE.pth nor models\Wan-AI\Wan2.1-T2V-1.3B\Wan2.1_VAE.pth found. Using models\Wan-AI\Wan2.1-T2V-1.3B\Wan2.1_VAE.pth as fallback. Loading models from: models\Wan-AI\Wan2.1-T2V-1.3B\diffusion_pytorch_model.safetensors Traceback (most recent call last): File "P:\AI\Wan21_v38\Wan2.1\venv\lib\site-packages\gradio\queueing.py", line 625, in process_events response = await route_utils.call_process_api( File "P:\AI\Wan21_v38\Wan2.1\venv\lib\site-packages\gradio\route_utils.py", line 322, in call_process_api output = await app.get_blocks().process_api( File "P:\AI\Wan21_v38\Wan2.1\venv\lib\site-packages\gradio\blocks.py", line 2103, in process_api result = await self.call_function( File "P:\AI\Wan21_v38\Wan2.1\venv\lib\site-packages\gradio\blocks.py", line 1650, in call_function prediction = await anyio.to_thread.run_sync( # type: ignore File "P:\AI\Wan21_v38\Wan2.1\venv\lib\site-packages\anyio\to_thread.py", line 56, in run_sync return await get_async_backend().run_sync_in_worker_thread( File "P:\AI\Wan21_v38\Wan2.1\venv\lib\site-packages\anyio\_backends\_asyncio.py", line 2470, in run_sync_in_worker_thread return await future File "P:\AI\Wan21_v38\Wan2.1\venv\lib\site-packages\anyio\_backends\_asyncio.py", line 967, in run result = context.run(func, *args) File "P:\AI\Wan21_v38\Wan2.1\venv\lib\site-packages\gradio\utils.py", line 890, in wrapper response = f(*args, **kwargs) File "P:\AI\Wan21_v38\Wan2.1\App.py", line 749, in generate_videos loaded_pipeline = load_wan_pipeline(model_choice, torch_dtype, vram_value, lora_path=effective_loras, lora_alpha=None) File "P:\AI\Wan21_v38\Wan2.1\App.py", line 1184, in load_wan_pipeline model_manager.load_models( File "P:\AI\Wan21_v38\Wan2.1\DiffSynth-Studio\diffsynth\models\model_manager.py", line 426, in load_models self.load_model(file_path, model_names, device=device, torch_dtype=torch_dtype) File "P:\AI\Wan21_v38\Wan2.1\DiffSynth-Studio\diffsynth\models\model_manager.py", line 408, in load_model if model_detector.match(file_path, state_dict): File "P:\AI\Wan21_v38\Wan2.1\DiffSynth-Studio\diffsynth\models\model_manager.py", line 165, in match if len(state_dict) == 0: TypeError: object of type 'NoneType' has no len()

cool1

Did you use the "windows_download_models.bat" file to download the models you want to use for Wan 2.1? though maybe they're separate to the models it's talking about there.

SECourses

Please download models you want to use

Hockey

So RIFE can make images between two images so they fit better together. I would love to see a feature where you can choose the first image and last image for a video you want to generate and then let RIFE fill in the blanks.

SECourses

Sadly RIFE is not that capable :D it generates next frame in a flowing frames. it is most capable of guessing next 2 frames.

绣 李

Traceback (most recent call last): File "D:\aiwords\Wan21_v38\Wan2.1\App.py", line 17, in import wan File "D:\aiwords\Wan21_v38\Wan2.1\wan\__init__.py", line 1, in from . import configs, distributed, modules File "D:\aiwords\Wan21_v38\Wan2.1\wan\modules\__init__.py", line 1, in from .attention import flash_attention File "D:\aiwords\Wan21_v38\Wan2.1\wan\modules\attention.py", line 11, in import flash_attn File "D:\aiwords\Wan21_v38\Wan2.1\venv\lib\site-packages\flash_attn\__init__.py", line 3, in from flash_attn.flash_attn_interface import ( File "D:\aiwords\Wan21_v38\Wan2.1\venv\lib\site-packages\flash_attn\flash_attn_interface.py", line 15, in import flash_attn_2_cuda as flash_attn_gpu ImportError: DLL load failed while importing flash_attn_2_cuda: 找不到指定的模块。

Walker4k

Can flux lora's be used with this?

SECourses

Nope you need Wan 2.1 LoRAs. Hopefully I will add 10s of LoRAs soon

SECourses

hi what is your GPU and do you have Python 3.10 installed?

绣 李

python3.10.11 cuda:12.4

绣 李

The model is downloading now, not sure if it's model related or not

SECourses

not models related. make a fresh install and send me install logs : monstermmorpg@gmail.com - you can move models folder to fresh install no worries

Han Vu

Hey Furkan, I get this error, while running the Batch Processing. It completes one video, but doesn't complete the RIFE improved version. Please help as soon as you can as I am just starting to run large batches: File "E:\Programs\Wan21_v40 - Copy\Wan2.1\App.py", line 1104, in batch_process_videos generation_details += f"CFG Scale: {cfg_scale}\n" NameError: name 'cfg_scale' is not defined

SECourses

i fixed this cfg error few minutes ago please update. your other error is more problematic. what is your GPU and python version?

Han Vu

OK. I will update. Sorry, I refreshed my window again and the single generation is running now.

Han Vu

Thank you Furkan, I will be testing the batch process again.

DanO..

I think you should have a poll asking what cards people have. Or at least how much VRAM. One for patrons and one for free visitors. You could find out which set up to focus on for your paying patrons and what might bring in free people that just visit. I'd suggest my card (of course) as one of the options RTX 3060 12G VRAM.

SECourses

it is really hard that way but i try to cover every GPU. that is why i added configs for every GPU. your GPU should work too but just slower if you have RAM.

DanO..

Oh yeah! I figure it out eventually. I just thought it might be good to know if you only had 2 people with 24+ Gigs of VRAM and most have much less. OR most people have 24 Gigs plus. Up to you.

Han Vu

I am reporting that the batch function is working fine now! Thank you. Also I have a question, can RTX 3090 do the FP8 (torch.float8_e4m3fn)? Which is recommended setting for RTX 3090. Thank you so much Furkan.

SECourses

sure FP8 will be faster since it uses lesser VRAM . select 24 GB and FP8 and it will auto set params

Han Vu

So based on your chart, I could do 6,000,000,000 params at FP8? What about 6,500,000,000?

Ash Murphy

Sorry if this may be a silly question. I have a computer with 4GB of VRAM. Does this mean I can still run the WAN I2V 720P 14B model on my computer via the MassedCompute setup that was outlined? Or does that relate to one of the smaller models?

Taiga

I was wondering if it's possible to add more video models.

Han Vu

Additionally FYI, FP8 selection so far hasn't worked twice. The video output is garbage. This is for RTX 3090. I tried with 24gb and also with 12gb selection.

SECourses

yes massed compute runs on cloud. you can even turn off your computer

SECourses

try same seed and prompt and steps and everything same with bf16 and see if makes difference

SECourses

currently configs are auto set between bf16 and fp8. they consider you are using maximum 2gb VRAM

Doey62750

Is it possible to make the first frame of the video and the last frame the same in order to make a loop?

Ash Murphy

I know that is possible on Luma - they have a loop or an extend feature. But would be good to know if it is achievable on the WAN model

SECourses

yes it is coming hopefully soon

Taiga

So, it's hardwired? I'll take a look at the code.

SECourses

yes i will do it. it will take last frame of generated video and generate a new video and stitch and go on

SECourses

yes currently it is supporting only Wan 2.1 loading

Philipp Capetian

If I want to update an existing installation in ubuntu, do I just extract the new version into the WAN Folder and rerun the installation? I did so but when launching it still said v42...

SECourses

No look at the Windows_Update_Version.bat file. you need to execute it as an .sh file. so it will git pull accurate remote repos. any LLM will give you sh file and command easy use it

Philipp Capetian

Is there an option to make looping or boomerang videos like in other models?

Taiga

sorry for the dumb question, but if I'm updating from v38, do I dump the entire contents of the v43 zip into it and run update? Also, can I move the install to another drive, or will it mess something up?

SECourses

you cant move installation. i suggest you to install v43 into new drive and move models folder. that will work best. updating from v38 - dump zip file and run updater should work

绣 李

RuntimeError: Cannot use ``weights_only=True`` with files saved in the legacy .tar format. In PyTorch 2.6, we changed the default value of the `weights_only` argument in `torch.load` from `False` to `True`. Re-running `torch.load` with `weights_only` set to `False` will likely succeed, but it can result in arbitrary code execution. Do it only if you got the file from a trusted source. Now it's this kind of mistake

绣 李

I was in China and spent three days downloading the model, which was very troublesome. I also hope that you can package and download the model. All installations were completed according to your requirements, but I still can't open it. Various problems are making me feel overwhelmed Traceback (most recent call last): File "D:\aiwords\Wan21_v38\Wan2.1\venv\lib\site-packages\gradio\queueing.py", line 625, in process_events response = await route_utils.call_process_api( File "D:\aiwords\Wan21_v38\Wan2.1\venv\lib\site-packages\gradio\route_utils.py", line 322, in call_process_api output = await app.get_blocks().process_api( File "D:\aiwords\Wan21_v38\Wan2.1\venv\lib\site-packages\gradio\blocks.py", line 2103, in process_api result = await self.call_function( File "D:\aiwords\Wan21_v38\Wan2.1\venv\lib\site-packages\gradio\blocks.py", line 1650, in call_function prediction = await anyio.to_thread.run_sync( # type: ignore File "D:\aiwords\Wan21_v38\Wan2.1\venv\lib\site-packages\anyio\to_thread.py", line 56, in run_sync return await get_async_backend().run_sync_in_worker_thread( File "D:\aiwords\Wan21_v38\Wan2.1\venv\lib\site-packages\anyio\_backends\_asyncio.py", line 2470, in run_sync_in_worker_thread return await future File "D:\aiwords\Wan21_v38\Wan2.1\venv\lib\site-packages\anyio\_backends\_asyncio.py", line 967, in run result = context.run(func, *args) File "D:\aiwords\Wan21_v38\Wan2.1\venv\lib\site-packages\gradio\utils.py", line 890, in wrapper response = f(*args, **kwargs) File "D:\aiwords\Wan21_v38\Wan2.1\App.py", line 748, in generate_videos loaded_pipeline = load_wan_pipeline(model_choice, torch_dtype, vram_value, lora_path=effective_loras, lora_alpha=None) File "D:\aiwords\Wan21_v38\Wan2.1\App.py", line 1312, in load_wan_pipeline model_manager.load_models( File "D:\aiwords\Wan21_v38\Wan2.1\DiffSynth-Studio\diffsynth\models\model_manager.py", line 426, in load_models self.load_model(file_path, model_names, device=device, torch_dtype=torch_dtype) File "D:\aiwords\Wan21_v38\Wan2.1\DiffSynth-Studio\diffsynth\models\model_manager.py", line 404, in load_model state_dict = load_state_dict(file_path) File "D:\aiwords\Wan21_v38\Wan2.1\DiffSynth-Studio\diffsynth\models\utils.py", line 69, in load_state_dict return load_state_dict_from_bin(file_path, torch_dtype=torch_dtype) File "D:\aiwords\Wan21_v38\Wan2.1\DiffSynth-Studio\diffsynth\models\utils.py", line 83, in load_state_dict_from_bin state_dict = torch.load(file_path, map_location="cpu", weights_only=True) File "D:\aiwords\Wan21_v38\Wan2.1\venv\lib\site-packages\torch\serialization.py", line 1541, in load return _legacy_load( File "D:\aiwords\Wan21_v38\Wan2.1\venv\lib\site-packages\torch\serialization.py", line 1787, in _legacy_load return legacy_load(f) File "D:\aiwords\Wan21_v38\Wan2.1\venv\lib\site-packages\torch\serialization.py", line 1654, in legacy_load raise RuntimeError( RuntimeError: Cannot use ``weights_only=True`` with files saved in the legacy .tar format. In PyTorch 2.6, we changed the default value of the `weights_only` argument in `torch.load` from `False` to `True`. Re-running `torch.load` with `weights_only` set to `False` will likely succeed, but it can result in arbitrary code execution. Do it only if you got the file from a trusted source.

SECourses

hi . make a fresh install and move models and it will work this is fixed. also i will upload models for manual download thanks for reminding

Floor13

Thanks for all the updates. After playing with ComfyUI for a couple of days (trying to recycle some old batch ones I made last year for images), I gave up and turned to your instead for practicing prompts. A couple of things. 1) Will your app have a product name? 2) I'm having an issue with the RIFE. The initial video generates, but then the RIFE step fails. This is Windows 11, RTX4090 24GB WAN 2.1 1.3B (Text/Video-to-Video) [CMD] Loading model: 1.3B with torch dtype: torch.float8_e4m3fn and num_persistent_param_in_dit: 7000000000 Loading models from: models\Wan-AI\Wan2.1-T2V-1.3B\diffusion_pytorch_model.safetensors model_name: wan_video_dit model_class: WanModel This model is initialized with extra kwargs: {'has_image_input': False, 'patch_size': [1, 2, 2], 'in_dim': 16, 'dim': 1536, 'ffn_dim': 8960, 'freq_dim': 256, 'text_dim': 4096, 'out_dim': 16, 'num_heads': 12, 'num_layers': 30, 'eps': 1e-06} The following models are loaded: ['wan_video_dit']. Loading models from: models\Wan-AI\Wan2.1-T2V-1.3B\models_t5_umt5-xxl-enc-bf16.pth model_name: wan_video_text_encoder model_class: WanTextEncoder The following models are loaded: ['wan_video_text_encoder']. Loading models from: models\Wan-AI\Wan2.1-T2V-1.3B\Wan2.1_VAE.pth model_name: wan_video_vae model_class: WanVideoVAE The following models are loaded: ['wan_video_vae']. Using wan_video_text_encoder from models\Wan-AI\Wan2.1-T2V-1.3B\models_t5_umt5-xxl-enc-bf16.pth. Using wan_video_dit from models\Wan-AI\Wan2.1-T2V-1.3B\diffusion_pytorch_model.safetensors. Using wan_video_vae from models\Wan-AI\Wan2.1-T2V-1.3B\Wan2.1_VAE.pth. No wan_video_image_encoder models available. num_persistent_val 7000000000 [CMD] Model loaded successfully. [CMD] Using resolution: width=832 height=480 100%|██████████████████████████████████████████████████████████████████████████████████| 50/50 [03:44<00:00, 4.49s/it] VAE decoding: 100%|██████████████████████████████████████████████████████████████████████| 9/9 [00:10<00:00, 1.22s/it] Saving video: 100%|███████████████████████████████████████████████████████████████████| 81/81 [00:00<00:00, 105.49it/s] [CMD] Applying Practical-RIFE with multiplier 2x FPS on video outputs\00002.mp4 Loaded 3.x/4.x HD model. Traceback (most recent call last): File "D:\Wan21_v43\Wan2.1\Practical-RIFE\inference_video.py", line 129, in lastframe = next(videogen) File "D:\Wan21_v43\Wan2.1\venv\lib\site-packages\skvideo\io\io.py", line 240, in vreader assert _HAS_FFMPEG, "Cannot find installation of ffmpeg." AssertionError: Cannot find installation of ffmpeg. Traceback (most recent call last): File "D:\Wan21_v43\Wan2.1\venv\lib\site-packages\gradio\queueing.py", line 625, in process_events response = await route_utils.call_process_api( File "D:\Wan21_v43\Wan2.1\venv\lib\site-packages\gradio\route_utils.py", line 322, in call_process_api output = await app.get_blocks().process_api( File "D:\Wan21_v43\Wan2.1\venv\lib\site-packages\gradio\blocks.py", line 2103, in process_api result = await self.call_function( File "D:\Wan21_v43\Wan2.1\venv\lib\site-packages\gradio\blocks.py", line 1650, in call_function prediction = await anyio.to_thread.run_sync( # type: ignore File "D:\Wan21_v43\Wan2.1\venv\lib\site-packages\anyio\to_thread.py", line 56, in run_sync return await get_async_backend().run_sync_in_worker_thread( File "D:\Wan21_v43\Wan2.1\venv\lib\site-packages\anyio\_backends\_asyncio.py", line 2470, in run_sync_in_worker_thread return await future File "D:\Wan21_v43\Wan2.1\venv\lib\site-packages\anyio\_backends\_asyncio.py", line 967, in run result = context.run(func, *args) File "D:\Wan21_v43\Wan2.1\venv\lib\site-packages\gradio\utils.py", line 890, in wrapper response = f(*args, **kwargs) File "D:\Wan21_v43\Wan2.1\App.py", line 907, in generate_videos subprocess.run(cmd, shell=True, check=True, env=os.environ) File "C:\Users\user\AppData\Local\Programs\Python\Python310\lib\subprocess.py", line 524, in run raise CalledProcessError(retcode, process.args, subprocess.CalledProcessError: Command '"D:\Wan21_v43\Wan2.1\venv\Scripts\python.exe" "Practical-RIFE/inference_video.py" --model="D:\Wan21_v43\Wan2.1\Practical-RIFE\train_log" --multi=2 --video="outputs\00002.mp4" --output="outputs\improved_00002.mp4"' returned non-zero exit status 1.

SECourses

hi. it is failing because your ffmpeg failing. please install exactly as in this video : https://youtu.be/DrhUHnYfwC0 and amazing updates coming to the app. today i am about to add a huge update extend video also i recommend you to install python as in video. your python is currently inaccurately installed. follow every step of this video : https://youtu.be/DrhUHnYfwC0

Floor13

Thanks for such a quick response. I spend so much time on PyPi at work, I hate messing with python when I get home. I had it all set up for AI generation 12 months ago, then got too busy with life and work so haven't been back. I left flash-attn building and the models downloading before I went to work, then found in your video that you'd already done the build :) I'll rip everything out and start again using the video guide you just provided tomorrow. Thank you, keep up the good work.

SECourses

awesome. that video will help you tremendously in long term

cool1

When you first load or reload the webpage (F5) that has the gradio app in it always sets the Lora field to the first lora it finds. If we don't want to use that Lora then we need to keep changing that Lora setting back to "None". Could the default value for the Lora setting be "None"?

SECourses

it sets to your last config lora. delete your configs and try and check

SECourses

yes i will publish batch RIFE processing App as soon as possible. also can you tell me how did you make it didnt work? i am testing right now that it works as expected

cool1

If the batch mode says "TypeError: WanVideoPipeline.__call__() got an unexpected keyword argument 'cancel_fn'" does that mean I need to do a full "windows_install.bat" or "windows_install_all_GPUs.bat" again? I'd done the "windows_update.bat" after extracting v45.zip

SECourses

delete DiffSynth-Studio and reinstall

cool1

Thanks. I tried that. I renamed it to DiffSynth-Studio-backup then ran Windows_install.bat and it's created a new diffsynth-studio folder but running batch mode still gives: File "G:\Wan-2.1-AI-Video-Generator-SEC-courses-ver-of-app\Wan2.1\App.py", line 1360, in batch_process_videos video_data = loaded_pipeline( File "G:\Wan-2.1-AI-Video-Generator-SEC-courses-ver-of-app\Wan2.1\venv\lib\site-packages\torch\utils\_contextlib.py", line 116, in decorate_context return func(*args, **kwargs) TypeError: WanVideoPipeline.__call__() got an unexpected keyword argument 'cancel_fn' I can try deleting everything and re-installing it. A previous time I'd used "windows install all gpus.bat". I don't know if that's what I should be using instead of just the "windows install.bat".

SECourses

i would say make a fresh install. i just tried image and video batch processing both works.

SECourses

hopefully even better than that training coming. i will prepare GUI as well

Philipp Capetian

Update v45 or v46? I did v45 today but did not notice

SECourses

last version has it. read the changelogs at the top.

Philipp Capetian

It states "Interface organization improved, many more updates coming soon hopefully". Or is there another log? Thanks anyways. I will give it a try tomorrow.

Ash Murphy

Amazing! I also have been having a little bit of trouble downloading the cudnn requirements to match your tutorial. It seems I don't get a .exe option to launch the installer, but instead get a bunch of .dll files. that have the likes of cudnn_ops_train64 or cudnn_adv_infer64 files in it. Is it likely I have done something wrong? I am pretty certain I went direct from the links on your tutorial page.

SECourses

you can use those dll files as well. extract into C drive and add to env path and will work. either way works

Walker4k

A small note, I have run these scripts locally (720p) on a 3090 and 5 sec render time takes a few (I recall 3 to 4) hours,. On Massedcompute, LS40, following instructions exactly as per YouTube video, about 90 mins, H100 around 40 mins. I also ran an H100 at Runpod, again no issues and similar times. For most home users or hobbyists, a $10 punt for a few hours on one of these machines is well worth it.

SECourses

Thanks. Also fast preset takes like 10 min on H100 at most demanding config 1280x720p image to video 81 frames

Weronika

i always get killed after click generate WAN 2.1 14B Image-to-Video 720P WAN 2.1 14B Image-to-Video 720P [CMD - DEBUG] Checking if pipeline needs clearing... [CMD - DEBUG] Pipeline config not changed or pipeline is None. No clearing needed. [CMD] Loading model: 14B_image_720p with torch dtype: torch.bfloat16 and num_persistent_param_in_dit: 22000000000 Loading models from: models/models_clip_open-clip-xlm-roberta-large-vit-huge-14.pth model_name: wan_video_image_encoder model_class: WanImageEncoder The following models are loaded: ['wan_video_image_encoder']. Loading models from: ['models/Wan-AI/Wan2.1-I2V-14B-720P/diffusion_pytorch_model-00001-of-00007.safetensors', 'models/Wan-AI/Wan2.1-I2V-14B-720P/diffusion_pytorch_model-00002-of-00007.safetensors', 'models/Wan-AI/Wan2.1-I2V-14B-720P/diffusion_pytorch_model-00003-of-00007.safetensors', 'models/Wan-AI/Wan2.1-I2V-14B-720P/diffusion_pytorch_model-00004-of-00007.safetensors', 'models/Wan-AI/Wan2.1-I2V-14B-720P/diffusion_pytorch_model-00005-of-00007.safetensors', 'models/Wan-AI/Wan2.1-I2V-14B-720P/diffusion_pytorch_model-00006-of-00007.safetensors', 'models/Wan-AI/Wan2.1-I2V-14B-720P/diffusion_pytorch_model-00007-of-00007.safetensors'] model_name: wan_video_dit model_class: WanModel This model is initialized with extra kwargs: {'has_image_input': True, 'patch_size': [1, 2, 2], 'in_dim': 36, 'dim': 5120, 'ffn_dim': 13824, 'freq_dim': 256, 'text_dim': 4096, 'out_dim': 16, 'num_heads': 40, 'num_layers': 40, 'eps': 1e-06} The following models are loaded: ['wan_video_dit']. Loading models from: models/models_t5_umt5-xxl-enc-bf16.pth model_name: wan_video_text_encoder model_class: WanTextEncoder The following models are loaded: ['wan_video_text_encoder']. Loading models from: models/Wan2.1_VAE.pth model_name: wan_video_vae model_class: WanVideoVAE The following models are loaded: ['wan_video_vae']. Using wan_video_text_encoder from models/models_t5_umt5-xxl-enc-bf16.pth. Using wan_video_dit from ['models/Wan-AI/Wan2.1-I2V-14B-720P/diffusion_pytorch_model-00001-of-00007.safetensors', 'models/Wan-AI/Wan2.1-I2V-14B-720P/diffusion_pytorch_model-00002-of-00007.safetensors', 'models/Wan-AI/Wan2.1-I2V-14B-720P/diffusion_pytorch_model-00003-of-00007.safetensors', 'models/Wan-AI/Wan2.1-I2V-14B-720P/diffusion_pytorch_model-00004-of-00007.safetensors', 'models/Wan-AI/Wan2.1-I2V-14B-720P/diffusion_pytorch_model-00005-of-00007.safetensors', 'models/Wan-AI/Wan2.1-I2V-14B-720P/diffusion_pytorch_model-00006-of-00007.safetensors', 'models/Wan-AI/Wan2.1-I2V-14B-720P/diffusion_pytorch_model-00007-of-00007.safetensors']. Using wan_video_vae from models/Wan2.1_VAE.pth. Using wan_video_image_encoder from models/models_clip_open-clip-xlm-roberta-large-vit-huge-14.pth. num_persistent_val 22000000000 [CMD] Model loaded successfully. [CMD] Using resolution: width=960 height=960 Killed (venv) Ubuntu@0066-dsm-prxmx30148:~/Desktop/wan/Wan21_v46/Wan2.1$

SECourses

you have to set virtual RAM. look how to set it. set it to 100 GB and it will get fixed

Floor13

Thanks for the new releases. I removed EVERYTHING (python, cudnn, etc) I had set up before, and started again from scratch following your video. Works perfectly THANK YOU! However: The new v46 version of your app, after running some 720p video generation, the Python process retains all the RAM (approx 29GB), while all the GPU RAM (24GB) is released when the generation is complete. Killing the app frees the python RAM. I did not see this with the v43 (But I had barely started using it). I am currently doing another set of multi-line, so I cannot provide clean logging at this time. (I see you've JUST updated to v47... I'll run through it after work).

Philipp Capetian

I updated to v47 but cannot see anything in the GUI that states loop or so...

SECourses

well i am trying to improve its RAM release. I release every object but still RAM management is not great. I wonder if related to Python itself at this moment.

Floor13

A question/comment/bug about seeds (I don't know if this is working as intended, or part on Wan2.1, or part of your UI). 1) I always set my own seed, just to make reproduction and trouble shooting easier. 2) For example: I set the seed to 0001 3) Image2video: I create similar prompts on 4 separate lines, to test a small change in each prompt, or an effect I want to try. 4) I check "Multi-line prompt" therefor 4 videos will be produced. 5) I would expect each video to have the same seed of 0001. 6) However, the seed increments by 1 each time, so vid 4 is 0004 seed. But I thought the seed should still be 0001. I am not running batches, I am not looking for variation, I am testing the prompt against the same seed. 7) Therefore, I am not testing the prompt alone, I am also introducing a new seed. Surely this is incorrect. If I wanted to build multiple vids with increasing seeds, then the "Number of generations" option should be used.

SECourses

thank you for detailed explanation. just updated to V50 and fixed this behavior. run updater. multi line will not increase seed anymore. and you can combine number of generations with multi line as well. in that case only number of generations will increase

Floor13

Wow! You work too hard! There was no rush for this, I have much bigger problems with my generations skill :)

abc clothing

Using the Wan2.1-I2V-14B-720P (Picture To Video) on a 4090RTX and 32GB RAM, everytime i get a bluescreen with memory error. Is there something i'm missing? Wan2.1-T2V-1.3B (Text To Video) works fine. I already tried reducing the vram preset.

SECourses

yes bluescreen means = your computer is not sable. are you using intel 13 14k series? i had 13900 and i hated it so much. it was constant blue screen shut down etc.

abc clothing

I'm using 13700kf, though I didn't think processor speed mattered as much for AI, like GPU and RAM do. Is there no chance i can get Wan2.1-I2V-14B-720P (PTV) to run? I don't think memory error BSOD relates to CPU either...

SECourses

you can run. but your RAM is not stable. you have to make it stable first. BSOD 100% related to your system instability. even if i use 200 GB RAM (i have 96GB rest is virtual) i don't get BSOD atm

Male AI Art

Very good. By the way, I'm using 3080ti to generate 480p image to video, is it normal that it takes about an hour?

Aixsponza

Doing the V2V feature I constantly receive this terminal error response: "RuntimeError: The size of tensor a (41) must match the size of tensor b (21) at non-singleton dimension 2" I can't find a setting to solve it.

SECourses

set your frame count accurately. what is your frame count?

Andrex Selivanov

I tried the application, and I want to say that it is very cool, although I had to manually install some packages, and the prompt enhancer did not work for me. Overall, its much more convenient than any I've used. thank you! did you tried it on 80gb(or more) runpod? how long it takes make 5 sec 720p on 80gb?

cool1

You could do a RAM/memory test in windows. If that's wrong it should say. You could check your GPU drivers are up to date. I've had occasional blue screen errors quite recently when trying to load too much into the GPU probably (like in GPT4ALL loading a model but the "gpu layers" were probably set too high so that might have been the cause of the blue screen error with that if it tried to load more into the GPU that it could hold. Or maybe Nvida's recent GPU drivers aren't very good or there some VRAM or RAM error (windows doesn't say any RAM error for me though).

SECourses

thank you so much. it takes lesser than 10 minutes. i just made some more improvements :d

SECourses

yes because of VRAM limitation. you can reduce steps and increase teacache to speed up

Aixsponza

Sorry, my stupid. I now set it to the clips frame count and all fine.

EvanGuo

能做个wan的lora训练教程吗

Male AI Art

Thank you for your reply, I reduced the number of steps and it greatly increased the generation speed. Let me ask one question, is it possible to use wan models published on civitai?

Philipp Capetian

I am still looking for an option to make looping videos. I mean with that that the last frame of a video matches with the first frame of the SAME video. I need short videos I can make animated gifs from for my powerpoint presentations. The new feature seems only good for extending videos - if I am not mistaken.

stphnvdb321

I have to report that i still have problems with videos longer than 10 seconds. I get heavy color changes and an over- or underexposed picture. It starts with video 3 or 4. Examples: video 1 https://postimg.cc/18NDMvBn, video 3 https://postimg.cc/gwV8brfG, video 4 https://postimg.cc/hJWTYVZC. I used 14B text-to-video, 30 steps, TeaCache 0.15. My VRAM was only half full. I've tried different settings like changing framerate, duration, aspect ratio, video quality. Nothing helped. Do you have any idea what can be the reason for this strange behaviour? Thanks.

SECourses

have you tested that last used frame and generated single video? i am gonna test it right now because i also see color changes. i wonder if related to our implementation or if it is image to video model issue

SECourses

i dont know understand what you mean can you elaborate? currently it works like this. generate a video, use last frame and extend, use last frame and extend and so on. or you can batch process folders too. or you can use existing video to extend as well

SECourses

yes it is coming hopefully soon as soon as i complete the lacking features in Wan 2.1 App

stphnvdb321

No, I haven't. I emptied the folder used_last_frames, because I thought the pictures could cause problems when I make a new video. So I can't do it right now. I found another issue: I increased steps from 30 to 50. With TeaCache enabled it produced heavy distortions beginning from 2.mp4

SECourses

yes i noticed same thing. 20 steps 81 frames looks really good with teacache. i am testing that i told you right now same parameters

Philipp Capetian

I want to create one single video (just a few seconds) that runs in a seamless loop. So the last frame and the first frame of the video should be the same. Otherwise it will be choppy when looping.

SECourses

ah i dont know how that can be made :D any ideas you have?

SECourses

perhaps there can be a lora that will make this. i will just request this

Philipp Capetian

I'm always using img2vid. So if wan could be instructed to use the same image as init and end frame, things would work out this way...

Raiga Xolotl

Will you ever cover Lora Training for WAN / HUN ? the Squishy Lora was awesome.

SECourses

after i complete this app, which we are close, hopefully i will make a full easiest and best tutorial for lora training

SECourses

well i tested and the app and it works. the degrade in quality gets increased since the last frames are getting worse with each extending :( but if your video is good it will work on some cases e.g. : https://streamable.com/of48pl

stphnvdb321

But there is definitly an improvement compared to the last version. Now it gets worse only in the end.

stphnvdb321

BTW: I just tested extended videos with "Number of Generations >1" and "Multi-line prompt (each line is separate)". It works perfectly! Thanks a lot for making this possible.

SECourses

awesome. it took me huge time to make it work :D i rewritten a lot of code

Ritesh 'Tez' B

found an issue i am having with v52. the second time i run this (I2V), it does not change the initial image. it works on the first generated image used from before. also noticec v52 did no do a auto pre processed image like the previous versions so i could check. reverted back to v49. it works again.

SECourses

can you give more info please i need to test and fix asap. step by step produce. also now we are v55

Ritesh 'Tez' B

It was a basic image to video prompt 480p. on Fast preset. 20INF, 5 seconds stadard. 3:4 aspect ratio. first attempt it changed a bit of the initial image. second attempt with another image ended up using the same image as the first one even though i am sure i changed it. i checked the auto pre processed image for initial frame. it had not saved the two i did in V52. Was a clean install but i moved the models, output, lora, auto preprossed images from v49 into the new V52.

SECourses

ok so you made a video with image to video. then you changed image uploaded new one but it used old image right? let me test

SECourses

thank you just updated v56 . please run update. it is fixed. ty for reporting

Bob Doyle Media

After installing, I get this error: Traceback (most recent call last): File "C:\Wan21_v55\Wan2.1\App.py", line 21, in import wan File "C:\Wan21_v55\Wan2.1\wan\__init__.py", line 1, in from . import configs, distributed, modules File "C:\Wan21_v55\Wan2.1\wan\configs\__init__.py", line 7, in from .wan_i2v_14B import i2v_14B File "C:\Wan21_v55\Wan2.1\wan\configs\wan_i2v_14B.py", line 3, in from easydict import EasyDict ModuleNotFoundError: No module named 'easydict' Press any key to continue . . .

SECourses

my mistake sorry for this error. please download v56 zip file and use it to install. it should fix. ty for reporting.

Bob Doyle Media

Thank you so much for your very prompt response. Installing now.

SECourses

that makes sense. Wan just announced yesterday a model for input image + ending frame coming. hopefully that can work as you want

SECourses

yes you can use Wan LoRAs published on there. also we just made another great speed update

stphnvdb321

Thanks for the new update. Sage really makes it faster. I made a 30-seconds video for fun. The second half was absolutely unusable and totally overexposed. Do you already know what could be causing this?

SECourses

yes because this is happening - as we progress the last frame quality degrades thus the output. it has a chain effect

stphnvdb321

A chain effect is bad :( I don't know how, but maybe you can find a way to fix it. I checked WAN 2.1 included in pinokio (Wan2GP). It uses Riflex to generate longer videos. Maybe this would be another way... https://github.com/deepbeepmeep/Wan2GP

Ivan Cruz

does not work after all the steps, the bat file (start app) opens and closes instantly :\

Grzegorz Busek

I really appreciate your work. I have 4080TI, but waiting 2 hours for 250KB output video using default settings reminds me of times when the Internet was shared on floppy disks :D (was using image to video 420p)

SECourses

haha true :D but it shouldnt take that long. especially 420p model. are you using teacache 0.15 and 20 steps? by the way hopefully GGUF coming very soon maybe this week

SECourses

that means you didnt follow requirements step by step no skip : https://youtu.be/DrhUHnYfwC0 also if you open cmd on that folder and type bat file full name it wont close and you can see error message

The90sPope

Ok, so the latest v56 version: I2V 480p, using fp8 setting, teacache default 0.15, sage attention, 81 frame, 20 steps. Running on 4080 16gb preset, got black screen video. Cfg was 6, shift was 10. Generated in 12mins. Any clue why black screen? Image I used was regular 832x480

The90sPope

Check windows error log, you are bsoding because your hardware is unstable. Most likely power delivery issue that can be countered with bios settings. If it's not power, then it's either gpu overclock or memory overclock that is unstable. Might be cpu overclock too. But I would check if it's power related first.

Larry Boles

Latest version fresh install, used the downloader to get the 1.3B model. Saved to: \models\Wan-AI\Wan2.1-T2V-1.3B when attempting to generate getting this error after the "Saving Video" stage: FileNotFoundError: No such file or directory: "models\\Wan-AI\\Wan2.1-I2V-14B-480P\\diffusion_pytorch_model-00001-of-00007.safetensors"

SECourses

yes for extending videos you need that model as well. so download it too

SECourses

these are all base models same as ours. i plan to add support to single model files and GGUF as well. hopefully coming soon

Larry Boles

Ah okay, I wasn't intending to run any video extension so I didn't get it. Probably should make a user note of that in the download_models batch file

Philipp Capetian

It generates a video and saves it. But it get's stuck then even when I press the cancel button...

stphnvdb321

I reduced Sigma Shift a bit and it helped a lot to reduce or even avoid exposure changes! I made some 20 seconds long videos and I had no changes of exposure at all or only a bit. 5.6 was a good value for me. Maybe the last frame was a bit to blurry with default setting 6? I don't know. Try it yourself.

SECourses

cancel , cancels the generation process and nothing else happens. what do you see on cmd? can you give me more info i will fix it hopefully asap

SECourses

you are right need to do . i am also working on some bug fixes

SECourses

yes those values can help. so you recommend default shift to be 6?

stphnvdb321

I prefer lower settings (the picture gets sharper). I tried also extended video with 5.0, but the attached video segments got ugly sharpening effects. I got the best results with 5.6. It's still not complety gone, but a lot better. Even a 50-second video was only overexposed at the end. Test it yourself to see the changes.

stphnvdb321

I see you've changed CFG Scale too. Does it have any effect on the video quality? I thought it only specifies how precise the model is executing your promt.

Philipp Capetian

[CMD] Loading model: 14B_image_480p with torch dtype: torch.bfloat16 and num_persistent_param_in_dit: 3500000000 Loading models from: models/models_clip_open-clip-xlm-roberta-large-vit-huge-14.pth model_name: wan_video_image_encoder model_class: WanImageEncoder The following models are loaded: ['wan_video_image_encoder']. Loading models from: ['models/Wan-AI/Wan2.1-I2V-14B-480P/diffusion_pytorch_model-00001-of-00007.safetensors', 'models/Wan-AI/Wan2.1-I2V-14B-480P/diffusion_pytorch_model-00002-of-00007.safetensors', 'models/Wan-AI/Wan2.1-I2V-14B-480P/diffusion_pytorch_model-00003-of-00007.safetensors', 'models/Wan-AI/Wan2.1-I2V-14B-480P/diffusion_pytorch_model-00004-of-00007.safetensors', 'models/Wan-AI/Wan2.1-I2V-14B-480P/diffusion_pytorch_model-00005-of-00007.safetensors', 'models/Wan-AI/Wan2.1-I2V-14B-480P/diffusion_pytorch_model-00006-of-00007.safetensors', 'models/Wan-AI/Wan2.1-I2V-14B-480P/diffusion_pytorch_model-00007-of-00007.safetensors'] model_name: wan_video_dit model_class: WanModel This model is initialized with extra kwargs: {'has_image_input': True, 'patch_size': [1, 2, 2], 'in_dim': 36, 'dim': 5120, 'ffn_dim': 13824, 'freq_dim': 256, 'text_dim': 4096, 'out_dim': 16, 'num_heads': 40, 'num_layers': 40, 'eps': 1e-06} The following models are loaded: ['wan_video_dit']. Loading models from: models/models_t5_umt5-xxl-enc-bf16.pth model_name: wan_video_text_encoder model_class: WanTextEncoder The following models are loaded: ['wan_video_text_encoder']. Loading models from: models/Wan2.1_VAE.pth model_name: wan_video_vae model_class: WanVideoVAE The following models are loaded: ['wan_video_vae']. Using wan_video_text_encoder from models/models_t5_umt5-xxl-enc-bf16.pth. Using wan_video_dit from ['models/Wan-AI/Wan2.1-I2V-14B-480P/diffusion_pytorch_model-00001-of-00007.safetensors', 'models/Wan-AI/Wan2.1-I2V-14B-480P/diffusion_pytorch_model-00002-of-00007.safetensors', 'models/Wan-AI/Wan2.1-I2V-14B-480P/diffusion_pytorch_model-00003-of-00007.safetensors', 'models/Wan-AI/Wan2.1-I2V-14B-480P/diffusion_pytorch_model-00004-of-00007.safetensors', 'models/Wan-AI/Wan2.1-I2V-14B-480P/diffusion_pytorch_model-00005-of-00007.safetensors', 'models/Wan-AI/Wan2.1-I2V-14B-480P/diffusion_pytorch_model-00006-of-00007.safetensors', 'models/Wan-AI/Wan2.1-I2V-14B-480P/diffusion_pytorch_model-00007-of-00007.safetensors']. Using wan_video_vae from models/Wan2.1_VAE.pth. Using wan_video_image_encoder from models/models_clip_open-clip-xlm-roberta-large-vit-huge-14.pth. num_persistent_val 3500000000 [CMD] Model loaded successfully. [CMD] Auto processed image saved to: auto_pre_processed_images/auto_processed_1742997965.png 100%|███████████████████████████████████████████| 50/50 [13:08<00:00, 15.78s/it] Saving video: 100%|█████████████████████████████| 41/41 [00:01<00:00, 26.90it/s] [CMD] Applying Practical-RIFE on original outputs/0003.mp4 Initializing RIFE model... done in 0.43s Press 'c' at any time to cancel the FPS increase process. Reading video metadata... done in 0.01s [INFO] Input: 41 frames at 8.00fps | Output: 32.00fps (4x) Setting up video reader... done in 0.52s [INFO] Audio will be transferred from source to output Setting up FFmpeg encoder... 0003_improved.mp4 [640x640@32fps] Setting up processing buffers... done Starting reader and writer threads... Preparing first frame... Pre-loading: Complete - 40 frames loaded in 0.04s (1102.7 fps) Processing frames: [------------------------------] 1/41 (2.4%) | 11008.7 fps | ETA: 0s Finalizing video output... done [SUCCESS] Processing complete! 40 frames in 2.53s (15.79 fps) Processing audio... No audio stream found in source video. Skipping audio transfer. [DONE] Completed video saved to: outputs/0003_improved.mp4 [CMD] Cancel button pressed. [CMD] Cancel button pressed.

Philipp Capetian

Sure. But rather uninformative. It is Done but the GUI is not responding but continues on and on processing. Hitting cancel is not stopping it. And it is not restarting if I start a new generation

Philipp Capetian

I analyzed the problem further with the help of ChatGPT: When checking running processes, I see the Python backend (App.py) and RIFE still running, even though processing is already complete. After force-killing the processes (kill -9), everything works again when I restart manually. Is this a known issue with WAN 2.1 or Practical-RIFE integration? Could it be related to the --share flag or resource limits? Any tips to avoid freezing or automatically reset the backend after processing?

SECourses

when RIFE processing we can't cancel true. cancellation only happens during steps stage.

Casper

For v2v input video needs to match exact frame count of the setting? I get tensor mismatch

SECourses

how many frames your video has and how many frames did you put as output? i can try to make auto cropped

stphnvdb321

Did you already check RIFLEx? https://github.com/thu-ml/RIFLEx. It should be able to expand the duration from 5 to 8 seconds (128 frames). And it is just one line of code. With your Extend Video Factor we would already have 16 seconds with factor 2, 24 seconds with factor 3 etc. It sounds very interesting to me.

SECourses

yes diffsynth authors said they will add still waiting

Walker4k

Error is Massedcompute - Using wan_video_vae from models/Wan2.1_VAE.pth. Using wan_video_image_encoder from models/models_clip_open-clip-xlm-roberta-large-vit-huge-14.pth. num_persistent_val 22000000000 [CMD] Model loaded successfully. [CMD] Auto processed image saved to: auto_pre_processed_images/auto_processed_1743239389.png 0%| | 0/20 [00:00

SECourses

hello i dont see error in your message

Rudbby

Gettting this error on massed computer: WAN 2.1 14B Text-to-Video [CMD - DEBUG] Checking if pipeline needs clearing... [CMD - DEBUG] Pipeline config not changed or pipeline is None. No clearing needed. 0%| | 0/50 [00:00 b s (n d)", n=num_heads) File "/home/Ubuntu/Desktop/best_ai/Wan2.1/venv/lib/python3.10/site-packages/einops/einops.py", line 600, in rearrange return reduce(tensor, pattern, reduction="rearrange", **axes_lengths) File "/home/Ubuntu/Desktop/best_ai/Wan2.1/venv/lib/python3.10/site-packages/einops/einops.py", line 532, in reduce return _apply_recipe( File "/home/Ubuntu/Desktop/best_ai/Wan2.1/venv/lib/python3.10/site-packages/einops/einops.py", line 251, in _apply_recipe tensor = backend.reshape(tensor, final_shapes) File "/home/Ubuntu/Desktop/best_ai/Wan2.1/venv/lib/python3.10/site-packages/einops/_backends.py", line 93, in reshape return x.reshape(shape) RuntimeError: CUDA error: no kernel image is available for execution on the device CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect. For debugging consider passing CUDA_LAUNCH_BLOCKING=1 Compile with `TORCH_USE_CUDA_DSA` to enable device-side assertions.

SECourses

hello where do you use? are you on your own linux? RTX 5000 series? use our latest runpod installer it supports. what is your GPU model?

Rudbby

I was on Massed Compute, I rebuilt on L40 instead of A100 and its working now. Thanks

Walker4k

Apologies - Massed Compute - H100 - Error below : File "/home/Ubuntu/Downloads/Wan21_v58/Wan2.1/DiffSynth-Studio/diffsynth/pipelines/wan_video.py", line 274, in __call__ noise_pred_posi = model_fn_wan_video(self.dit, latents, timestep=timestep, **prompt_emb_posi, **image_emb, **extra_input, **tea_cache_posi) File "/home/Ubuntu/Downloads/Wan21_v58/Wan2.1/DiffSynth-Studio/diffsynth/pipelines/wan_video.py", line 385, in model_fn_wan_video x = block(x, context, t_mod, freqs) File "/home/Ubuntu/Downloads/Wan21_v58/Wan2.1/venv/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1751, in _wrapped_call_impl return self._call_impl(*args, **kwargs) File "/home/Ubuntu/Downloads/Wan21_v58/Wan2.1/venv/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1762, in _call_impl return forward_call(*args, **kwargs) File "/home/Ubuntu/Downloads/Wan21_v58/Wan2.1/DiffSynth-Studio/diffsynth/models/wan_video_dit.py", line 208, in forward x = x + gate_msa * self.self_attn(input_x, freqs) File "/home/Ubuntu/Downloads/Wan21_v58/Wan2.1/venv/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1751, in _wrapped_call_impl return self._call_impl(*args, **kwargs) File "/home/Ubuntu/Downloads/Wan21_v58/Wan2.1/venv/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1762, in _call_impl return forward_call(*args, **kwargs) File "/home/Ubuntu/Downloads/Wan21_v58/Wan2.1/DiffSynth-Studio/diffsynth/models/wan_video_dit.py", line 143, in forward x = self.attn(q, k, v) File "/home/Ubuntu/Downloads/Wan21_v58/Wan2.1/venv/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1751, in _wrapped_call_impl return self._call_impl(*args, **kwargs) File "/home/Ubuntu/Downloads/Wan21_v58/Wan2.1/venv/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1762, in _call_impl return forward_call(*args, **kwargs) File "/home/Ubuntu/Downloads/Wan21_v58/Wan2.1/DiffSynth-Studio/diffsynth/models/wan_video_dit.py", line 117, in forward x = flash_attention(q=q, k=k, v=v, num_heads=self.num_heads) File "/home/Ubuntu/Downloads/Wan21_v58/Wan2.1/DiffSynth-Studio/diffsynth/models/wan_video_dit.py", line 50, in flash_attention x = sageattn(q, k, v) File "/home/Ubuntu/Downloads/Wan21_v58/Wan2.1/venv/lib/python3.10/site-packages/sageattention/core.py", line 134, in sageattn return sageattn_qk_int8_pv_fp8_cuda_sm90(q, k, v, tensor_layout=tensor_layout, is_causal=is_causal, sm_scale=sm_scale, return_lse=return_lse, pv_accum_dtype="fp32+fp32") File "/home/Ubuntu/Downloads/Wan21_v58/Wan2.1/venv/lib/python3.10/site-packages/torch/_dynamo/eval_frame.py", line 838, in _fn return fn(*args, **kwargs) File "/home/Ubuntu/Downloads/Wan21_v58/Wan2.1/venv/lib/python3.10/site-packages/sageattention/core.py", line 829, in sageattn_qk_int8_pv_fp8_cuda_sm90 assert SM90_ENABLED, "SM90 kernel is not available. Make sure you GPUs with compute capability 9.0." AssertionError: SM90 kernel is not available. Make sure you GPUs with compute capability 9.0.

SECourses

i was expecting it to be supported by all GPUs. this is bad.

SECourses

thanks new installer will work on all gpus. hopefully i will compile for each gpu later

SECourses

thanks new installer will work on all gpus. hopefully i will compile for each gpu later

SECourses

hello. use new GPU it should fix. sorry for this error

Male AI Art

v59 I double clicked on Download_models.py. What is happening? File "E:\Wan21_v59\Download_Models.py", line 171, in main() File "E:\Wan21_v59\Download_Models.py", line 159, in main download_model_in_two_phases(repo_id, model_folder, common_items) File "E:\Wan21_v59\Download_Models.py", line 22, in download_model_in_two_phases snapshot_download(repo_id=repo_id, local_dir=model_specific_dir, ignore_patterns=common_items) File "C:\Users\****\AppData\Local\Programs\Python\Python310\lib\site-packages\huggingface_hub\utils\_validators.py", line 120, in _inner_fn return fn(*args, **kwargs) File "C:\Users\****\AppData\Local\Programs\Python\Python310\lib\site-packages\huggingface_hub\_snapshot_download.py", line 237, in snapshot_download _inner_hf_hub_download(file) File "C:\Users\****\AppData\Local\Programs\Python\Python310\lib\site-packages\huggingface_hub\_snapshot_download.py", line 215, in _inner_hf_hub_download return hf_hub_download( File "C:\Users\****\AppData\Local\Programs\Python\Python310\lib\site-packages\huggingface_hub\utils\_validators.py", line 120, in _inner_fn return fn(*args, **kwargs) File "C:\Users\****\AppData\Local\Programs\Python\Python310\lib\site-packages\huggingface_hub\file_download.py", line 1291, in hf_hub_download raise LocalEntryNotFoundError( huggingface_hub.utils._errors.LocalEntryNotFoundError: Connection error, and we cannot find the requested files in the disk cache. Please try again or make sure your Internet connection is on.

Villi Manter

Hey friend, I really need your help. I've spent so much time trying to fix this issue, but nothing has worked so far 😔 I'm using Gradio 4.15.0 with LivePortrait (custom gr.Blocks UI), and I keep getting this error: APIInfoParseError: Expected dict, got : True It crashes the whole app and I can't even load the interface — just a 500 error on launch. I tried editing get_type() in gradio_client/utils.py, but still no luck. Here’s the full traceback: https://drive.google.com/file/d/1iU6kyjgH8vt2oLRCGSVP8jcXDUHRAMFQ/view?usp=sharing Please let me know what I'm missing 🙏

Cloud MaxGaming

Hello my friend if you could make a version of( InfiniteYou: Flexible Photo Recrafting While Preserving Your Identity) to work on mass computing also if you could review infiniteYou

SECourses

hello. it is on my todo list sorry for delay

Antonio

Hi, I have a laptop with 4090 and 16GB GPU RAM. When I try to make a image-to-video the exceution time is about 40 minutes. Could you tell me the best configuration to decrease the execution time?

Antonio

for text-to-video is about 1 hour and 30 minutes

SECourses

16 GB GPU + apply fast preset are fastest. have you tested

SECourses

16 GB GPU + apply fast preset are fastest. have you tested

Antonio

yes, my best execution time was 26 minutes with I2V 480P, but quality is not good.

Antonio

sorry but always about 30 minutes with many settings combination

SECourses

thank you for reporting. i am expecting GGUF support to added to our library this week. hopefully it will reduce VRAM requirement significantly to get better quality at low VRAM GPUs

Rizwan Sheikh

I trained a t2v Lora on musubi-tuner and it works with that package's inference code. When trying to use this, I'm stuck on a non-descriptive "Cannot load LoRA: LoRAs\loraname.safetensors" Any ideas of what I could check or do to figure out why? Also, it may be beneficial to others if it doesn't proceed with generating anyway if it can't load a specified LoRA; it took me a while to realize it was not loading

Walker4k

Is it possible to add start and end image for the I2V workflow?

SECourses

Wan team said they will publish that model very soon probably this week. hopefully i will add

SECourses

hello. can you send me lora file so we can analyze and fix hopefully. you can email me link : monstermmorpg@gmail.com

Lautaro Javier

Image to video doesn't work for me, but text to video does. My GPU is a 3090. When I click generate, the command window crashes without displaying any error messages.

Dinho Medeiros

Sir, Im sorry if this was asked before, but I coudnt find any information about this: Is there any way of setting this up to make video requests through api? It would be a game changer for me. If you have any tutorials about this, let me know please.

Marc Mercuri

I didnt see guidance on utilizing batch processing in the video. If I have multiple prompts and each prompt tied to a different input image, how is that set up?

Larry Boles

Would love to see if we can use VACE with this in the future

man

Amazing program as always! I used the fast preset for the 14b img2vid model and is it normal for the program to take 4-5 minutes before it even shows the progress (it/s) bar? It seems like the program doesn't even start for a long time! Using 4070ti Super (16gb VRAM)

Steve

Which file and line contains the path to the model folder in windows? I would like to edit it to a different folder. Thank you

Steve

Hello I am getting this error. I have already deleted .trition cache and tried reinstalling and still getting the error Thanks! import error: DLL load failed while importing libtriton: A dynamic link library (DLL) initialization routine failed. W0405 16:36:07.336000 5468 venv\Lib\site-packages\torch\distributed\elastic\multiprocessing\redirects.py:29] NOTE: Redirects are currently not supported in Windows or MacOs. Traceback (most recent call last): File "C:\WanGradio\Wan21_v60\Wan2.1\App.py", line 25, in from diffsynth import ModelManager, WanVideoPipeline, save_video, VideoData File "C:\WanGradio\Wan21_v60\Wan2.1\DiffSynth-Studio\diffsynth\__init__.py", line 2, in from .models import * File "C:\WanGradio\Wan21_v60\Wan2.1\DiffSynth-Studio\diffsynth\models\__init__.py", line 1, in from .model_manager import * File "C:\WanGradio\Wan21_v60\Wan2.1\DiffSynth-Studio\diffsynth\models\model_manager.py", line 4, in from .downloader import download_models, download_customized_models, Preset_model_id, Preset_model_website File "C:\WanGradio\Wan21_v60\Wan2.1\DiffSynth-Studio\diffsynth\models\downloader.py", line 6, in from ..configs.model_config import preset_models_on_huggingface, preset_models_on_modelscope, Preset_model_id File "C:\WanGradio\Wan21_v60\Wan2.1\DiffSynth-Studio\diffsynth\configs\model_config.py", line 57, in from ..models.wan_video_dit import WanModel File "C:\WanGradio\Wan21_v60\Wan2.1\DiffSynth-Studio\diffsynth\models\wan_video_dit.py", line 21, in from sageattention import sageattn File "C:\WanGradio\Wan21_v60\Wan2.1\venv\lib\site-packages\sageattention\__init__.py", line 1, in from .core import sageattn, sageattn_varlen File "C:\WanGradio\Wan21_v60\Wan2.1\venv\lib\site-packages\sageattention\core.py", line 20, in from .triton.quant_per_block import per_block_int8 as per_block_int8_triton File "C:\WanGradio\Wan21_v60\Wan2.1\venv\lib\site-packages\sageattention\triton\quant_per_block.py", line 18, in import triton File "C:\WanGradio\Wan21_v60\Wan2.1\venv\lib\site-packages\triton\__init__.py", line 8, in from .runtime import ( File "C:\WanGradio\Wan21_v60\Wan2.1\venv\lib\site-packages\triton\runtime\__init__.py", line 1, in from .autotuner import (Autotuner, Config, Heuristics, autotune, heuristics) File "C:\WanGradio\Wan21_v60\Wan2.1\venv\lib\site-packages\triton\runtime\autotuner.py", line 9, in from .jit import KernelInterface File "C:\WanGradio\Wan21_v60\Wan2.1\venv\lib\site-packages\triton\runtime\jit.py", line 12, in from ..runtime.driver import driver File "C:\WanGradio\Wan21_v60\Wan2.1\venv\lib\site-packages\triton\runtime\driver.py", line 1, in from ..backends import backends File "C:\WanGradio\Wan21_v60\Wan2.1\venv\lib\site-packages\triton\backends\__init__.py", line 50, in backends = _discover_backends() File "C:\WanGradio\Wan21_v60\Wan2.1\venv\lib\site-packages\triton\backends\__init__.py", line 43, in _discover_backends compiler = _load_module(name, os.path.join(root, name, 'compiler.py')) File "C:\WanGradio\Wan21_v60\Wan2.1\venv\lib\site-packages\triton\backends\__init__.py", line 12, in _load_module spec.loader.exec_module(module) File "C:\WanGradio\Wan21_v60\Wan2.1\venv\lib\site-packages\triton\backends\amd\compiler.py", line 2, in from triton._C.libtriton import ir, passes, llvm, amd ImportError: DLL load failed while importing libtriton: A dynamic link library (DLL) initialization routine failed.

SECourses

models folder has it. you may symlink i presume or change from app code directly Wan2.1\models

SECourses

it could be loading model. how much RAM you have? what Disk you are using your disk speed?

SECourses

sorry for late reploy. imagename.jpg png etc and in same folder imagename.txt , it will use that txt as a prompt

SECourses

you can start gradio and every gradio has api feature. but we have no other api feature

SECourses

please tell me the error you are getting. you can message me from discord

Steve

I have installed build tools 2022 LTSC 17.8.17 with all options and deleted other versions and reinstalled the app, but I am getting the same error. Any ideas?

Steve

Hello again, I installed the build tools per the instructions in your video but am still getting the error after reinstallation of your app. Any ideas?

SECourses

what is the error message? if same as before install visual studio and MSVC as well. and delete triton cache again if you upgrade gold membership i can connect your pc and try to fix

Steve

I installed visual studio pro 2022 V17.13.5 on top of the build tools LTSC and the app now launches. Thanks! I get this error at startup but it still works: W0407 19:48:07.706000 25996 venv\Lib\site-packages\torch\distributed\elastic\multiprocessing\redirects.py:29] NOTE: Redirects are currently not supported in Windows or MacOs.

Steve

Great app, especially with VRAM management! For the first time, I was able to generate larger sized videos than with comfy. Feature Request: Multi-LoRa support please! Also, a generation preview would be nice! (getting these errors, but does not affect function:) At startup: W0407 21:12:11.389000 24652 venv\Lib\site-packages\torch\distributed\elastic\multiprocessing\redirects.py:29] NOTE: Redirects are currently not supported in Windows or MacOs. After generation: \AppData\Local\Programs\Python\Python310\lib\asyncio\proactor_events.py", line 165, in _call_connection_lost self._sock.shutdown(socket.SHUT_RDWR) ConnectionResetError: [WinError 10054] An existing connection was forcibly closed by the remote host

SECourses

we have multi lora already :D click Show More LoRAs that error you get at the end is related to Gradio and your PC i don't know exact solution but not important

Steve

Oh, nice!

Steve

Hello, what is causing this error for V2V? I already tried resizing and matching the framerate. File "C:\WanGradio\Wan21_v60\Wan2.1\DiffSynth-Studio\diffsynth\schedulers\flow_match.py", line 67, in add_noise sample = (1 - sigma) * original_samples + sigma * noise RuntimeError: The size of tensor a (30) must match the size of tensor b (21) at non-singleton dimension 2 Edit: I did not match the framerate correctly. You must first match using an external video editor. Works at 16 fps, not sure about other frame rates.

SECourses

latest app version does this auto. did you update the app?

绣 李

After installing, I get this error: Traceback (most recent call last): File "C:\Wan21_v55\Wan2.1\App.py", line 21, in import wan File "C:\Wan21_v55\Wan2.1\wan\__init__.py", line 1, in from . import configs, distributed, modules File "C:\Wan21_v55\Wan2.1\wan\configs\__init__.py", line 7, in from .wan_i2v_14B import i2v_14B File "C:\Wan21_v55\Wan2.1\wan\configs\wan_i2v_14B.py", line 3, in from easydict import EasyDict ModuleNotFoundError: No module named 'easydict' Press any key to continue . . .

Leonid Wolf

Is it possible to use your platform remotely? like with Swarm through cloudflare?

SECourses

yes. edit the Start_APP.bat and make it like python App.py --share and save thats it

SECourses

hello. install failed. get the latest zip file and use it please. and you can email me install logs : monstermmorpg@gmail.com

Leonid Wolf

Can you tell me how to prescribe it? I don't quite understand how to do this.

Steve

I think the original generation mp4 file may have a codec issue. Some video players cannot play the file eg, android video player or windows media player.

Tom

Hello, great course, love your work. I have a very strange issue though. I ran the Wan2.1-I2V-14B-480p model yesterday quite successfully and generated 10 or so videos in a reasonable speed (took about an hour for each video), running on a 4080 Super. But today, I went to do the same thing, and it takes *infinitely* longer now - it took over two hours just to load the model, and it's been stuck on 0% progress now for several more hours. Something has clearly changed, but I am using the exact same model and settings as yesterday. I have not installed or modified anything on my PC since yesterday. I know this is a vague problem, but has anyone encountered this type of issue before?

Floor13

I had the same issue last night after I updated from v60 to v62. (After about 30 mins at 0% I killed it.) After restarting the app and the UI, it worked again. Assuming it is all the memory that is retained by Python, or some issue where it checks to see if the same version is loaded, finds it is loaded and tries to work, but maybe some issue there... Of course I keep forgetting that logs are not saved.

DanO..

Really looking forward to new features! I have an RTX 3060 12GB and can do 14B 480 videos overnight. Fun. If I try to add a LoRA python crashes. Is this to be expected with my sort of card? Or might there be something I can change to get LoRAs to work?

SECourses

we didnt do any changes since yesterday. perhaps something else using your VRAM or RAM or disk?

SECourses

LoRA is not depended on GPU. it is using totally same VRAM and GPU power. please set your virtual RAM to 100 GB and it won't crash. for 12 GB hopefully I am gonna add GGUF as soon as possible. sorry for this delay

SECourses

could be. it is from DiffSynth-Studio . but RIFE improved should be playable.

DanO..

Thanks buddy! I appreciate you.

SECourses

it is so easy. locate Start_APP.bat. edit it with any text editor like notepad.

SECourses

awesome. i get that error too ignore it

100Sachen80

Hi Furkan, how can i generate a "preview" video with small quality but very fast, so that i can then decide if the output is good for what i promped and then generate the exactly same video but with better quality? should i keep inference steps, but lower quality to 1 or 2, teacache to 0.3 for example, and resolution much lower for "preview" video? which values can i change and then use higher of these values for consistent output (same frames, same video) as final video?

Leonid Wolf

"share" is not an internal or external command, executable program, or batch file.

SECourses

sadly we dont have it atm. i asked this to diffsynth library but they didnt add yet

Leonid Wolf

Thanks! We look forward to your assembly with the gguf models)))

100Sachen80

But i could do that manually? Lower Res, lower quality, steps 20. faster generation of video. check result. if i like the result. increase Res, increase quality, steps 20. final video. but it must mean that the fast video is the same as the final video regarding the content. same seed of course.

Ahmet Inceelli

Happy to hear that GGUF support is coming soon.. 😸

Ryan game

Hello! New member here 8^] I'm using the 14B i2v model on a 4090 and it's taking 136 seconds per it, any idea why it could be taking so long? Windows 11, AMD Ryzen Threadripper 2920X , 64gb RAM, RTX 4090

SECourses

hello. are you using 1280x720p? it is really long. you can apply fast preset . hopefully I will add GGUF support this week and welcome. you can reduce resolution it really speeds up or reduce number of frames it also speeds up

Ryan game

No that was the 480 model. Ended up taking 38 minutes for 81 frames @ 480x720 which is weid because the same generation takes about 5.5 minutes using comfyui... but I 'm reinstalling right now, I'm not sure if Sagattention installed properly. I don't think i had all of the requiremets installed. I'll send you a update when I try again! I appreciate the fast response!

Ryan game

Is there a way to tell if sageattention is installed/ working properly?

Tom

Thanks for the replies, all. I couldn't figure out what was causing it, I tried restarting my PC/app, closing every single background process beforehand - nothing worked. But then the next day, it went back to working normally again. It's still working fine now. Very strange issue, but at least it's resolved now! It must have been some issue with my machine. Thanks again for the amazing app.

SECourses

it doesnt print which attention but can you tell me your step speed? that sounds too slow. i got rtx 5090 and it is way faster and 4090 not that much slower. open nvitop and tell me how much power your GPU is using. if it is using not almost max, that means it is using shared VRAM and you need to select lower VRAM preset

SECourses

great thanks for letting me know

Alexandr Mishuk

how do you get i2v 720p model generation in 4-5 minutes I tried on my 3090 - 30+ minutes on l40 - same about 20-30 minutes a100 - took about 1000 seconds - 16 minutes

SECourses

are you doing fast preset? RTX 5090 and H100 can get below 10 min

Necati Kılınçarslan

Thank you very much. Thanks to you, I finally found a working solution for my rtx 5090 system. I want to use this feature for this workflow start end frame (transition between two photos). can the feature be added?

SECourses

that feature will come soon hopefully not yet. and thanks and welcome

Alexandr Mishuk

"are you doing fast preset?" - Yes - Uh, so 10 minutes is the best you can do. It's becoming uneconomical.

SECourses

yep not great speed atm. 480p plus upscaling can work. i am working on adding the most powerful AI video upscale process

Léo Ulr

Hello, First, thank you very much for your work ! Controlnet with this would be awsome ! I hope you will implement it one day :) In the meantime, I need a small advice: I try to generate a simple videos with i2v, with 8fps and 81 frames (10 seconds). Resolution 720x1280 (I have RTX 4090). A character that is simply dancing on the spot, I want the hands and arms to stay static (because they go crazy most of the time !). But at the end of the videos, the hands and arms don't seem to follow the prompt, and everything is messed up. The first seconds of videos are perfect. I tried different prompt (and negatives) and I tried to play with the sigma shift parameter without success. But it seem it don't follow the prompt in the lasts frames. Any advices ? Being able to choose the first frame as last frame would be very usefull (loop)! I'll continue experimenting :)

SECourses

thanks. well hopefully I will add new model soon that has beginning and ending frames. that may fix your issue. other than that i am not sure. did you try lower sigma shift and cfg values?

Pablo Montero

Ey Furkan, I would like to check with you if my generation times are normal. Image-to-Video WAN 2.1 14B Image-to-Video 720P, default parameters, around 42 minutes 121.92s/it with a RTX 3090. What do you think?

100Sachen80

Hi Furkan, the zip file points to v64 instead of new v66

Ariloum

Is there any reasoning why did you prefer unsafe pth (which allows to inject malicious code) over safetensors? For example, downloading models script leads to models_t5_umt5-xxl-enc-bf16.pth while there is long time exists umt5-xxl-enc-bf16.safetensors.

SECourses

i used whatever the official authors used :D not my preference

SECourses

it installs latest from remote repo so zip file update necessary only when installers updated

SECourses

sounds right. that model is huge 720p

Ariloum

I'm trying to manage models and folders with symbolic links for comfy/forge/LLM's with all these heavy files (I already got terabytes of them on fastest SSD's) and I have scripts for automating that needed folders symbolic linkage, but builds like this trying to reinvent the wheel downloading same models in weird/compromised formats (like splitting by pieces Wan2.1 safetensors model or this text encoding in pth) to strange folders structures and it's harder to manage model files this way.. How did you manage fat model files library by yourself? Did you have just tens of copies of same models for many different projects?

SECourses

i just purchased 2x 4 TB NVME :D

Ariloum

yeah I have 3x4TB and 1x8TB SSD's but still trying to avoid having same models spreading across many endpoints

SECourses

wow you got nice. makes sense to prevent duplicate

stphnvdb321

Still no RifleX? This would be a cool feature to generate longer videos. Some other suggestions: Pytorch compilation to increase speed (on Linux or using WSL on Windows), Skip Layer Guidance ( should improve video quality) and Classifier-Free Guidance Zero Star.

SECourses

i have some things to implement. i am also tracking them

stphnvdb321

I just saw them in another Wan21 app and these are really cool features. I know you're busy... Thanks for all your work!

TAGopolis

Anyone know where to add a --listen command to allow this to be used on my network?

SECourses

edit the app.py file with text editor and go very bottom and change like below demo.launch(share=args.share, inbrowser=True, server_name="0.0.0.0")

Ariloum

this installation is not working to me, I run install-all bat, then update bat, and models download. then if I run it I got error "no library" for easydict and diffusers, I got installed them manually. And then it shows error "module gradio has no attribute Blocks" File "C:\stable-diffusion\Wan21\Wan2.1\App.py", line 2278, in with gr.Blocks() as demo: AttributeError: module 'gradio' has no attribute 'Blocks'......

Ariloum

nvm I got it fixed (reinstall helps), also I messed up with models folder structure...

Guck Foogle

if you get the ERROR: THESE PACKAGES DO NOT MATCH THE HASHES FROM THE REQUIREMENTS FILE, its because the pip install torch command in the script doesnt download the same nightly build as the command, (who knows why) i had to download manually the 3 whl from https://download.pytorch.org/whl/nightly/cu128/ and replace the pip that install torch by pip install torch-2.7.0.dev20250310+cu128-cp310-cp310-win_amd64.whl torchvision-0.22.0.dev20250311+cu128-cp310-cp310-win_amd64.whl torchaudio-2.6.0.dev20250311+cu128-cp310-cp310-win_amd64.whl

SECourses

thanks. i also made a fix in zip file

David

I've been getting this error with V67 and V65v2 (Worked fine before) in massedcompute: bash: venv/bin/activate: No such file or directory /home/Ubuntu/.pyenv/versions/3.10.16/bin/python3: can't open file '/home/Ubuntu/Downloads/Wan21_v65_v2/Wan2.1/App.py': [Errno 2] No such file or directory Ubuntu@0155-dsm-prxmx30154:~/Downloads/Wan21_v65_v2/Wan2.1$ I also noticed that there's 2 wan 2.1 subfolders: "wan 2.1" and "Wan 2.1" (capitalized) don't know if this helps.

SECourses

yes folder name issue. fixing now

SECourses

ty for reporting fixed with v66 zip file. i am also testing now on a fresh pod

David

It seems to be working now! Thanks!

Jay

Any hope of the start frame end frame implementation? I believe they finally released it

SECourses

yes hopefully coming very soon. sorry for delay currently working on framepack. it is also amazing and has even time based prompt you tested?

Jay

No worries Doc! Was just wondering :) and no I haven’t tried it yet!

SECourses

ye give it a try it is mind blowing so good. supports begin + end with timed prompts. I am generating a 30 second video :D

tes

It seems that in the Batch Process, the output file names used to be the same as the file names in the Input Folder with an .mp4 extension, but in recent versions, they are now numbered sequentially starting from 0001. Is this a bug, or has the specification changed? Personally, I find it more convenient if the output files have the same names as those in the Input Folder.

Erkel Ermint

Hi, I need your help to find out what went wrong. All pre-requisites were installed successfully. After pressing the "Generate" button, the following error msg stopped the generation: ValueError: Unsupported CUDA architecture: sm61 I have a GTX1080 with 8GB VRAM. 2 versions of CUDA are installed, v11.8 and v12.4. Tried to change the default by modifying the env variables. But no success.. :(

SECourses

1080 is just too old. you cant run this. but you can try framepack. download install and then use Uninstall_Sage_Attention_For_Better_Quality_Slower.bat file and try. it may work

MLGuy

Hi , I'm encountering some issues when using 14B models. The models do load successfully, and my graphics card appears to handle the initial processing without any problem. However, right before the output is generated, the connection times out and the session closes on its own. Additionally, the I2V (Image-to-Video) functionality doesn't seem to work at all. The only model that works consistently for me is WaN 2.1 (1.3B) in T2V mode. I would really appreciate it if you could provide some detailed guidance or troubleshooting steps to help resolve these issues. Thank you!

SECourses

hi please set your virtual ram to at least 50 GB. gradio connection error wouldnt prevent output to be saved. i think you are getting out of RAM error with big models

MLGuy

Here are my Specs- 9800X3D,RTX 5090, 32 GB ram 6000Mhz CL30. Do You think i have to change some settings in Gradio for low ram usage?

SECourses

sadly Gradio doesnt have RAM setting only has VRAM setting. your GPU is top tier but only 32 GB Ram is very very low for that GPU. I recommend to upgrade your ram to at least 64 GB but preferably over 96 and more. so currently set your virtual ram to at least 50 GB preferably like 100 GB

MLGuy

Sure,I would try it .Also i have DRAM NVME 2TB SSD,Would that help? Thanks for ur help.

SECourses

definitely set at least 50 GB virtual RAM on that disk

SECourses

this was indeed a bug . sorry for fixing late. i just made a fix. can you test thoroughly and let me know if any other bugs remains? update to v67. just run update bat file

MLGuy

Hey, I got it working — enabled virtual memory up to 80GB to get past the VRAM limits. The model loads fine, but performance-wise it doesn't feel significantly better than the 1.4B version. On my RTX 5090, it still took around 25 minutes for a single run, which feels a bit underwhelming given the hardware. I was expecting more noticeable improvements in quality or efficiency. Are there any recommended ways to optimize prompts for better coherence or output richness with this model? Also, are there flags or config tweaks (like quantization or reduced attention layers) that could help speed things up without sacrificing too much quality? I'd appreciate any suggestions to make the most out of this setup

SECourses

i think better prompting but not much i know atm

Todd E

What version of ffmpeg are you using? I followed the tutorial and downloaded a recent release but I am getting an invalid arg exception when running practical-rife. If I turn off rife, everything works fine. ffmpeg is in my PATH, I just think the version I'm on is expecting arguments it's not being given. Thank you!

SECourses

what is the error you are getting? my ffmpeg is : ffmpeg version N-118369-g959b799c8d-20250127 Copyright (c) 2000-2025 the FFmpeg developers also delete your RIFE folder and run installer again and see if it fixes. use latest zip file

Todd E

Such a fast response! I'm on `n7.1.1-6-g48c0f071d4-20250503`and the exact error I get is: Traceback (most recent call last): File "C:\AI\Wan21_v68\Wan2.1\Practical-RIFE\inference_video.py", line 220, in clear_write_buffer ffmpeg_proc.stdin.write(frame_bgr.tobytes()) OSError: [Errno 22] Invalid argument I'll downgrade my ffmpeg and see if that works, and if not I'll fall back and run the installer again. I'll also let you know what resolved it so you can further assist the community. Thank you!

SECourses

wow afterall it was ffmpeg error. i will link specific ffmpeg to installer post thanks for letting me know

Chro Cut

I've been getting error, there doesn't seem to be any App.py in the indicated folder. Error is: C:\Python310\python.exe: can't open file 'E:\\Wan21_v68\\Wan2.1\\App.py': [Errno 2] No such file or directory Press any key to continue . . .

SECourses

hi that means your install failed. do you have git installed? are you running as administrator? can you make a fresh install and send me cmd logs : monstermmorpg@gmail.com

Chro Cut

Tried re-installing and it worked, thanks! and keep up the good work.

DanO..

I updated last night from version 68. Now loras won't load and I get the error Error Loading Lora, "only integer tensors of a single element can be converted to an index". Is this something that can be fixed? Or do I need to go back? Or do you think something else is wrong?

SECourses

delete venv , DiffSynth-Studio folders and use latest zip file and run installer again. i have no issues lora working

DanO..

OK. Thanks.

BurneR BlasT CasT

Start app.bat file not starting app. It pops up then disappears

SECourses

you must have install error. open a cmd in that folder and type bat file name to see error

BurneR BlasT CasT

Wan21_v68\Wan2.1>call .\venv\Scripts\activate.bat || exit /b The system cannot find the path specified.

BurneR BlasT CasT

I've tried reinstalling and its the same issue, How do I resolve this?

SECourses

your git clone failed. could be vpn you are using or your antivirus or you didnt follow requirements : https://youtu.be/DrhUHnYfwC0 - make a fresh install and send me cmd logs : monstermmorpg@gmail.com

Jony Das

Start_APP not working, How can I do ? Everything I Install in my computer but Start-APP is not working!

abbker8899

Install worked and models loaded, Python 3.10.11, but when I hit start I get, ImportError: DLL load failed while importing libtriton: A dynamic link library (DLL) initialization routine failed. Any thoughts? BTW, got it working on another computer. Amazing program! Thanks!

Eclipse Beast - TF

In the Windows_Download_Models.bat file, you don't start the virtual environment, so I need to have Hugging Face installed in my clean environment. I've added the environment initialization myself, but just out of curiosity—was there a specific reason you chose not to initialize it in Windows_Download_Models.bat?

Dazzastrous

Your behind on these updates there is Vace 2.1 out now....:)

SECourses

true but it is not easy to catch up :D

SECourses

well make sure to have installed c++ tools, MSVC, CUDA 12.8 and clear triton cache.

SECourses

well i need more info. if you make a fresh install and send me logs i can know : monstermmorpg@gmail.com - send me entire cmd screen of fresh install

abbker8899

That did it. Thank you!

DarkView AI

stuck on building wheel for flash attn for 20 mins on l40s

BrainFartArt

works AWESOME! i use your Forge and now the WAN2.1 in addition to comfy and i managed to recreate the Carchase scene (5sec) in 3 min Render time on a 5090 https://civitai.com/models/1620122/car-chase-wan21-i2v-lora?modelVersionId=1833536 now experimenting with the CausVid Lora from ComfyUi in your WAN UI - seems to work good so far with cfg 1 and 35 steps (+ car chase lora and the image as reference) now i stole a few cool images off civitai and try to animate them - what a blast!

SECourses

thanks a lot for great feedback

Heisenberg

Hi, thank you for this. Sorry if this was answered before. If it was, would you consider add it to the post FAQs? I'm not sure how you would go to load a new model not supported in your downloader. E.g: https://civitai.green/models/1295569/on-the-fly-wan-ai-wan21-video-model-multi-specs-causvidandcomfyandkijai-workflow-included onTHEFLYWanAIWan21VideoModel_kijaiWan21I2V14B480P.safetensors I've checked the WAN2.1 models folder in the app but they usually contains a lot of other files and pieces, while this is a huge single file labeled as checkpoint merge. How can we use this model to speed up production? I don't think you'd load it as a LoRA, right? If not, how can I add it to the app to be listed with the other supported models. Thanks!

SECourses

hello. latest installer installs pre compiled wheel so it shouldnt compile itself

Heisenberg

Oh, thank you, I have that installed as well, will give it a try. Thanks!

togan

@SECourses Using latest installer Wan21_V68_2 VM Massed Compute. I found the WAN 14B i2v output poor quality. The bitrate data rate of video output 5 second video 480p resolution was barely over 1000kbps and only 1Mb in size. Compare to previous installer v68 the bitrate was much higher 4000+ and larger file size 2 to 3Mb much better quality. Same settings generate.

SECourses

set the quality parameter to 10. i just did it and made it default

Firat Güven

Is there any difference between wan21_v68 setup and wan21_v68_2 except the new lora model? Do I need to reinstall from v68 to v68_2? Most probably no, sorry if I my question has been already answered before.

SECourses

nope you dont need to just download the lora model and use

100Sachen80

Can we use fast preset for causvid in you gradio app, or better leave the defaults as they are?

Fernando

Wait... NO VACE IMPLEMENTATION??? Then why bother? This is June 2025, Video-2-Video is the ONLY way to go. Okay, canceling my membership again jeezzz....

SECourses

for VACE i will cover another tutorial hopefully soon

Gus

Why not just ask about it or see if it can be implemented before whining and throwing a fit? Or implement it yourself, you're paying 5 dollars a month or 16 cents a day for him to build things that you clearly can't. I'll pay your membership if you stop whining in the comments.

Fernando

All the stuff in this channel is copied left and right, practically nothing is made by this creator. This issue is not the 5 dollars, it's the constant advertising of the second coming. If you don't understand why v2v is the only relevant tech than you're the whiner, thinking you're adding value by supporting this channel, so leave the conversation to the grown-ups.

Fernando

Oh get this, "gus" isn't supporting ANYONE on any channel. Talking a big game but being the ultimate WHINER. Put your money where your mouth is, leecher.

Gus

Are you okay? Are you sure you're not just blocked from seeing who I support? Lol. Not going to spend a lot of time if this because I can already tell that you're not the type of person who isn't going to stop whining to the world about small things that effect them. So I will just say, if this guy is just ripping stuff off left and right why don't you do the same and make it the way you want? Or go outside and get some fresh air, it's not this serious.

Fernando

Kid, I'm sharing advice directly from the industry field. Where as you are adding NOTHING, proving again with your meaningless reply. Yes, I might have formulated it more kind, but I was disappointed with another useless tool, should have checked it better. It's already established YOU'RE the whiner, because whiners add NOTHING. Say something meaningful, I dare you. But you're just this typical unhappy person that needs to lash out at others to feel anything positive. But you won't. Because you are NOTHING.

Fernando

And by the way, don't use the "are you okay" or "do you need a hug" replies... that REALLY shows how desperate you are to "win" an "argument", which of course never happens, as you are still deeply unhappy with yourself, and nothing can cure that, because you still need to MAKE something of yourself.

koon grizzly

Installed this a couple of times already, my problems are still the same, even with 3090 it takes way longer then framepack and quality is so bad i can't even see what was created.. Check the example, all i get is things like this (picture to video 480p version) https://youtu.be/g8v3tdsbXjI for the first part, i had it set to extend but fell a sleep and woke up to this : https://www.youtube.com/watch?v=oXxU-YWI6vw, it gets a little better at the end but still looks afwul :(

koon grizzly

wow, just WOW.... Please cancel your membership so we don't have to read this annoying behavior of a child that didn't get his candy anymore

Fernando

RTFM buddy, already canceled. And I got the candy from someone else who DOES understand AI video developments. So good luck being stuck in the past.

Fernando

You don't support anyone on Patreon, so again, why the big mouth boy? Why don't you quit as it's clear you don't know what you're doing, whining about bad quality pffff good Lord.

koon grizzly

hahahahahahhahaha boehoehoehoehoe, i'm so unhappy now because some 6 year old baby called me names on the internet woehoehoehoe hahahahahahahhah, On a serious note : WHO THE F* do you even think you are talking to people like that if you know nothing about them ... And WHY are you still here if it's all so bad.. The person from this patreon makes one click installers that have everything included and save hours of time + he has tutorials that come with it that explain everything good, even for people who first hear about ai so he does a great job in what he does. If that is not what you are looking for then you are at the wrong place and attacking the wrong people, you should look in the mirror and blame yourself instead of playing the victim and attacking strangers on the internet

Gus

I mean you have to see the irony in this comment right?

Fernando

Haha grizzly and Gus are the same person right? Because you ran out of "arguments", right? And how pathetic it is you took time on THE WEEKEND to write something like that hahaha!! You're only here because you want to make adult content, so you're disgusting. But YOU commenting OUT OF THE BLUE on someone who complains about a paid subscription, calling it "whining" (look it up in the dictionary, that's not at all what I was doing, I was a tad unfriendly, that's all), making things personal et cetera et cetera, totally unprovoked... YOU are the PROBLEM you child. I have "won" this "argument" before it even started. You're pathetic and a total loser. "Both" of you hahaha. END OF "DISCUSSION". For your future "retorts", reread this comment.

Gus

Are you a real person? You seem like an AI bot that was promted to respond like a child that didn't get enough attention at home, and is very "edgy".

James P

Can you add queuing like you have in framepack?

SECourses

not atm sadly. but we have batch processing. also open same url in second tab set and hit generate and gradio will auto queue. this works with all gradio apps

Stevie Tee

Hi do you have any plans to implement the new Wan 2.1 FusionX models into this gradio version , I'm struggling to get them to work in ComfyUI, even trying the GGUF models.

SECourses

i am about to publish a tutorial for it recorded

Stevie Tee

Excellent ! I shall look forward to it.

SECourses

sorry for delay i am editing still

Acid Hunter

Hi! first of all thank you for your tutorials are amazing! I got a very low end laptop, well as good specs but a very limited video card in this case an T1200 nvidia 4gb vram is a workstation. The thing is when I ran Wan2.1 gets to the UI but then no matter what I do gets an error. when I check the cmd shows ValueError: Unsupported CUDA architecture: sm75, this is somethign I can fix with drivers or it's the Wan model that not support this card? thanks in advance

SECourses

sadly you have no chance to run. your best case is using massed compute or runpod. i recommend massed compute

Acid Hunter

I see, I am trying to put to the max a medium end laptop to generate some content with AI, probably will upgrade to a better gpu or use an external as a workaround. Thanks for your response! btw two other questions do you have a list of models for generating images that could work fine with 4gb vram i am not looking to run fast due my limitations but can generate good quality images? and second could you make a tutorial using n8n to call local generator images? Thank you again!

Eon Kalpa

I have problem running Windows_Download_Models.bat. The error is [INFO] Downloading selected model(s)... Traceback (most recent call last): File "D:\Wan21_v68_2\Download_Models.py", line 4, in from huggingface_hub import snapshot_download ModuleNotFoundError: No module named 'huggingface_hub' Please help me!

SECourses

for 4 gb vram you can generate images but very slow with SwarmUI

Eon Kalpa

Help! I have installed and started the app. I'm doing WAN 2.1 14B i2V 480p with torch.float8_e3m3fn. I'm using RTX5090 card with 32GB VRAM. Generating just 5sec videos from an image. All the videos generated are just black screens. Tried 50 inference steps and 20 inference steps. Quality=5. What am I doing wrong?

Eon Kalpa

Ok just to update, I managed to get it working now. Basically everything has to be installed on C: drive. Previously I had installed some components on C: drive and the app installed on D: drive. Now after I changed everything to C: it works fine.

Ritan Mohajan

File "/home/Ubuntu/Downloads/Wan21_v68_2/Wan2.1/DiffSynth-Studio/diffsynth/models/model_manager.py", line 423, in load_model if model_detector.match(file_path, state_dict): File "/home/Ubuntu/Downloads/Wan21_v68_2/Wan2.1/DiffSynth-Studio/diffsynth/models/model_manager.py", line 167, in match if len(state_dict) == 0: TypeError: object of type 'NoneType' has no len() I am getting this error during video generation. Also I was getting following when 2nd command had been executed. "Folder 'xlm-roberta-large' not found in Wan2.1/models/Wan-AI/Wan2.1-T2V-14B; skipping copy." Enter your choice (1-6): 4 Phase 1: Downloading specific files for Wan-AI/Wan2.1-T2V-14B from IntervitensInc/Wan2.1-T2V-14B-FP16 ... .gitattributes: 2.35kB [00:00, 8.53MB/s] LICENSE.txt: 11.4kB [00:00, 43.0MB/s] README.md: 15.6kB [00:00, 63.7MB/s] .DS_Store: 100%|███████████████████████████| 6.15k/6.15k [00:00<00:00, 42.6MB/s] assets/comp_effic.png: 100%|███████████████| 1.60M/1.60M [00:00<00:00, 3.07MB/s] assets/data_for_diff_stage.jpg: 100%|████████| 528k/528k [00:00<00:00, 1.88MB/s] assets/i2v_res.png: 100%|████████████████████| 892k/892k [00:00<00:00, 3.24MB/s] assets/input.png: 100%|████████████████████| 1.07M/1.07M [00:00<00:00, 3.81MB/s] logo.png: 100%|████████████████████████████| 56.3k/56.3k [00:00<00:00, 2.81MB/s] assets/t2v_res.jpg: 100%|████████████████████| 301k/301k [00:00<00:00, 1.19MB/s] assets/vben_1.3b_vs_sota.png: 100%|██████████| 516k/516k [00:00<00:00, 1.97MB/s] assets/vben_vs_sota.png: 100%|███████████████| 584k/584k [00:00<00:00, 1.67MB/s] assets/video_dit_arch.jpg: 100%|█████████████| 643k/643k [00:00<00:00, 1.73MB/s] assets/video_vae_res.jpg: 100%|███████████████| 212k/212k [00:00<00:00, 612kB/s] config.json: 100%|█████████████████████████████| 250/250 [00:00<00:00, 2.67MB/s] diffusion_pytorch_model-00001-of-00006.s(…): 100%|█| 4.94G/4.94G [00:09<00:00, 5 diffusion_pytorch_model-00002-of-00006.s(…): 100%|█| 4.92G/4.92G [00:07<00:00, 6 diffusion_pytorch_model-00003-of-00006.s(…): 100%|█| 4.92G/4.92G [00:10<00:00, 4 diffusion_pytorch_model-00004-of-00006.s(…): 100%|█| 4.92G/4.92G [00:10<00:00, 4 diffusion_pytorch_model-00005-of-00006.s(…): 100%|█| 4.92G/4.92G [00:10<00:00, 4 diffusion_pytorch_model-00006-of-00006.s(…): 100%|█| 3.96G/3.96G [00:08<00:00, 4 (…)ion_pytorch_model.safetensors.index.json: 96.8kB [00:00, 149MB/s] examples/i2v_input.JPG: 100%|█████████████████| 251k/251k [00:00<00:00, 742kB/s] special_tokens_map.json: 6.62kB [00:00, 27.3MB/s] google/umt5-xxl/spiece.model: 100%|████████| 4.55M/4.55M [00:00<00:00, 9.88MB/s] google/umt5-xxl/tokenizer.json: 100%|██████| 16.8M/16.8M [00:00<00:00, 33.7MB/s] tokenizer_config.json: 61.7kB [00:00, 148MB/s] Phase 1 complete for Wan-AI/Wan2.1-T2V-14B. Phase 2: Downloading common files for Wan-AI/Wan2.1-T2V-14B from IntervitensInc/Wan2.1-T2V-14B-FP16 ... Wan2.1_VAE.pth: 100%|█████████████████████████| 508M/508M [00:01<00:00, 310MB/s] models_t5_umt5-xxl-enc-bf16.pth: 100%|██████| 11.4G/11.4G [00:35<00:00, 322MB/s] Phase 2 complete for Wan-AI/Wan2.1-T2V-14B. Copied folder 'google' from Wan2.1/models/Wan-AI/Wan2.1-T2V-14B to Wan2.1/models. Folder 'xlm-roberta-large' not found in Wan2.1/models/Wan-AI/Wan2.1-T2V-14B; skipping copy. Press Enter to exit...

SECourses

hi model downloads failed. please redownload model and pay attention. make sure that your machine has enough free disk space also use Wan21_v68_2.zip

SECourses

yes i understood. it can still happen. so when you get such error run downloader again. also make sure has sufficient disk on the machine

koon grizzly

i meanwhile found out what was the problem after watching the tutorial once again, there is a setting of quality that is at 10 when i install the software. In your tutorial at some moment you mention default is 5, so i changed this and now it works very well. At least for 5 second videos, then it looks very good, after that it degrades very fast

koon grizzly

i am surprised he still came back and cry for more attention from strangers on the internet, maybe it indeed is a bot because that not how real people (who are old enough to use a computer) react.

lp

Hello, I'm testing your interface for the first time, but I’ve encountered several blocking issues — some of which I’ve managed to partially fix. 1. Web interface not loading (blank page on launch) When launching the tool using the provided .bat file, the browser opened a blank (white) page and nothing was rendered. What I did to fix it: I manually edited the .bat script to: navigate to the correct folder containing app.py, activate the virtual environment (venv\Scripts\activate.bat), and launch app.py directly. After this change, the interface loads properly in the browser. 2. Gradio version issue I was initially using Gradio 5.36.2, which caused UI or loading issues when launching the app. Solution: I downgraded to an older, compatible version of Gradio (3.x), which resolved the UI problems and allowed the interface to display correctly. 3. CLIP model .pth fails to load Once the interface works, triggering a generation results in an error related to loading the CLIP model. Model file used: models_clip_open-clip-xlm-roberta-large-vit-huge-14.pth The file exists, It’s correctly named, File size is about 4.6 GB, I’ve placed it in all relevant model folders used by the app. Error message: The model is empty or unreadable (state_dict is empty or missing when using torch.load()). My questions: Is this .pth file compatible with your tool? If not, is there an official download link for a verified version of the model? Which exact version of Gradio do you recommend for full compatibility? I'm running this on a machine with an NVIDIA RTX 5090, so GPU compatibility should not be an issue. Thanks in advance for your help!

SECourses

it is gradio error. i just fixed it in remote repo. other error is download model related. i am making a fresh install again to verify . a model download might have failed for any reason

lp

Hi, just wanted to thank you! Everything is working perfectly now with the Gradio fix. I really appreciate your support and quick updates. 🙏

Cacoia Mecoia

Hello, I have been using Stable Diffusion for a while, and now I just installed WAN2.1 with environment described in this post, but running Start_APP.bat shows me the message: Traceback (most recent call last): File "C:\Users\user\Downloads\Wan21_v68_2\Wan2.1\App.py", line 26, in import wan File "C:\Users\user\Downloads\Wan21_v68_2\Wan2.1\wan\__init__.py", line 1, in from . import configs, distributed, modules File "C:\Users\user\Downloads\Wan21_v68_2\Wan2.1\wan\configs\__init__.py", line 7, in from .wan_i2v_14B import i2v_14B File "C:\Users\user\Downloads\Wan21_v68_2\Wan2.1\wan\configs\wan_i2v_14B.py", line 3, in from easydict import EasyDict ModuleNotFoundError: No module named 'easydict' Press any key to continue . . . I already tried pip install easydict, which seems to be ok: Microsoft Windows [Version 10.0.22631.5624] (c) Microsoft Corporation. All rights reserved. C:\Users\user>pip install easydict Requirement already satisfied: easydict in c:\python310\lib\site-packages (1.13) Please help. Thanks! Gabriel

100Sachen80

Hi Furkan, just to clarify: Wan FusionX (no separate file for 480/720p) is the best method for image to video for now? so no need for older tutorial with base wan 2.1 model, or wan 2.1 and phantom lora or causvid lora? or is phantom lora best in combination with base model? could you do a ranking?

100Sachen80

Yes, find it a bit confusing with all the options and loras in combination with wan or wan fusionx. in your latest tutorial you did not use any lora but just fusionx model, right? or maybe i missed it in the video.

SECourses

i have shown both i think. we have presets that makes it easier

Cacoia Mecoia

Any thoughts how to fix easydict?

SECourses

easydict means you had an error during install. please make a fresh install and send me install logs completey so i can see : monstermmorpg@gmail.com

Cacoia Mecoia

Hi, reinstalled and worked, thank you very much!

Alexandr Mishuk

Will WAN 2.1 FusionX be installed in this software, or is it only available in Swarm UI?

SECourses

currently only available in swarmui. we may add to this later. i am currently focused on video upscaler app

Alexandr Mishuk

Well, at least please make a ComfyUI setup 🙏 I'm really not a fan of Swarm UI — just can't get used to it. It would be awesome to have WAN 2.1 FusionX available in ComfyUI too!

Alexandr Mishuk

I only see workflow for multitalk. but I didn't find any for fusionx/sel force or any fast model :(

Robert K

Do you have any guides to extend the length of generated videos. Like maybe load a 5 second wan 2.2 video, generate another 5 seconds and stitch the two videos together?

SECourses

this app has it but for wan 2.2 swarmui has a feature which i havent tested yet. it is called as Video Extend. enable Display Advanced Options and look main tabs

Blockdeek

Hi mate, I literally had a thrilling moment after 50/50 [3:59:42<00:00, 287.65 s/it] said, "Error Connection error occurred." \o/ what could cause this ? No any log available. All UI Windows show Error

SECourses

hi it still did save it. this is gradio error. just checkout the outputs folder

Blockdeek

no just the tmp file > 0011.mp4.tmp - I ran 14b ItV 720p with an 3090. I will try again with a 1 second lenght to check whether it works. Will keep you updated

SECourses

oh it could be out of RAM too so process killed. did you set virtual RAM?

Sqit

Awesome tuorial. However I got strange error when launch app: Traceback (most recent call last): File "I:\wan2\Wan2.1\App.py", line 24, in import cv2 File "I:\wan2\Wan2.1\venv\lib\site-packages\cv2\__init__.py", line 181, in bootstrap() File "I:\wan2\Wan2.1\venv\lib\site-packages\cv2\__init__.py", line 153, in bootstrap native_module = importlib.import_module("cv2") File "C:\Users\asste\AppData\Local\Programs\Python\Python310\lib\importlib\__init__.py", line 126, in import_module return _bootstrap._gcd_import(name[level:], package, level) ImportError: DLL load failed while importing cv2:

SECourses

sorry for this error. get Wan21_v72.zip extract all files and overwrite previous files. enter inside wan 2.1 folder delete venv folder. then return back and then run Windows_Install_Or_Update.bat . this should fix your issues

Sqit

Found the problem. I have N version of windows. Install of media pack end restart solved the problem. (Some DLL was missing?)

SECourses

nice finding i dont know sadly :D

Myche Barrett

Could this be used to create ambiance videos from a photo? More specifically, I could give it a photo, tell it that I'd like to add say, rain to it, motion, etc...? Can it also add background sounds? Like say, ambient street noise or rain? If not this app, then which one would you recommend for this sort of thing?

Myche Barrett

Awesome! Ill read up on them and install these. Thank you for getting back with me and for all that you do...