Home Creators Posts Import Register Favorites Logout

Downloads

Content

FramePack, lllyasviel's image-to-video model, makes 30 FPS videos up to 120 seconds long in an advanced Gradio app with Sage Attention and Teacache support.

1-click installers for Windows, RunPod and Massed Compute run with as low as 6 GB VRAM and support RTX 3000 to 5000 series and cloud GPUs. V63 moved to uv installs, and the robust model downloader hash-verifies and resumes downloads; LoRA, batch processing and metadata features are included.

1-Click installers and super advanced Gradio App for lllyasviel FramePack image-to-video model. It can generate 30 FPS 120 second videos

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)

=======

Latest installer zip file : FramePack_v63.zip

Tutorial video : https://youtu.be/HwMngohRmHg

  • It uses Sage Attention and supports Teacache

  • Follow requirements very carefully

  • Supports RTX 5000 series as well

  • This is an optimized image to video model that leverages Hunyuan and models are automatically downloaded into hf_download folder inside main FramePack install folder

  • Official repo : https://github.com/lllyasviel/FramePack

9 March 2026 Updates V63

  • Install errors fixed

  • Installers upgraded to uv install thus it will be like 100x faster on RunPod and 10x faster on Windows and Massed Compute

16 October 2025 Updates V61

  • Installer files for RunPod Massed Compute and Windows updated

  • Now we moved to latest CUDA 12.9 and Torch 2.8 and Flash Attention 2.8.3 and Sage Attention 2.2.0 and xFormers 0.0.33

  • Now we support all GPUs very well starting from RTX 3000 series to 5000 series and all cloud GPUs like RTX A6000, A100, H100, B200 etc

  • For older GPUs than 3000, please message me from discord so i might have a solution for you this time

  • I also have completey re-written model download system and moved it to our newest ultra robust downloader

  • So it will 100% accurately download, hash verify and now it has full resume capability

  • If your download fails for any reason just run Windows_Download_Resume_Fix_Models.bat and it will resume

  • Upgrade to this version from V60 not needed but if you want, you need to do a fresh install since model load from disk system changed

  • We also have alternative FramePack Studio Installer now : https://www.patreon.com/posts/128472621

Windows Requirements

  • Python 3.10.11, FFmpeg, CUDA 12.9 or above, cuDNN 9.4 or above, C++ tools, MSVC and Git

  • If it doesn't work make sure to below tutorial and install everything exactly as shown in this below tutorial

  • https://youtu.be/DrhUHnYfwC0

27 May 2025 Updates V59

  • Since now there is a FramePack LoRA to test, I fixed our LoRA loading method and now LoRAs working

  • You can download and use this LoRA from here : https://civitai.com/models/1247109/high-speed-drone-shot-hunyuan-video-lora

    • I used activation prompt : low altitude high speed drone shot

    • So example : a car going fast on a road low altitude high speed drone shot

  • Just Windows_Update.bat should be sufficient

22 May 2025 Updates V57

  • Generate Metadata Immediately checkbox implemented after requested

  • It is enabled by default

  • It will generate a temp metadata as soon as process starts and make it saved as final file name after video has been fully generated

  • GPU Inference Preserved Memory (GB) description updated

    • Increase this until you don't use shared VRAM. If you increase unnecessarily it will become slower. However If you make this lower than expected, It will use shared VRAM and will become ultra slow. Follow nvitop and monitor GPU watt usage

  • Batch processing bug fixed

  • When batch processing and auto cropping / downscaling, it will take into consideration exif metadata of image for correct orientation now

  • Just run windows update bat file and it should be sufficient

17 May 2025 Updates V53

  • Since requested multiple LoRA option implemented

  • Check your GPU watt usage and if it is too low, that means it is using shared VRAM, that means increase GPU Inference Preserved Memory

  • This is rather a bit experimental you may install into a different folder and test

  • Just run windows update bat file and it should be sufficient

14 May 2025 Updates V52

  • Auto resize to nearest bucket of selected resolution feature implemented for batch processing

  • So lets say you did set 480 output resolution, 480 means 480x480=230,400 pixels. The program will downscale the input image into that total pixel count with keeping aspect ratio and then minimally crop to make width and height divisible to 32.

    • So for example 2560x3299 pixel image becomes Resolution: 480, Final Width: 416, Final Height: 544

  • Just run windows update bat file and it should be sufficient

12 May 2025 Updates V49

  • Few batch processing errors fixed

  • Now you can generate videos from text (basically it generates a black image with your desired target resolution - this is how others doing too)

  • Target resolution system implemented, now you can set any resolution you want - width and height has to be divisible to 32

  • When you upload an image, it will auto set target width and height according to the selected older Resolution and most appropriate bucket but you can change it if you want

8 May 2025 Updates V47

  • Load metadata feature implemented

  • Now you can drag and drop and load saved metadata to apply used settings

  • Just run windows update bat file

6 May 2025 Updates V46

  • FramePack F1 implemented more info here : https://github.com/lllyasviel/FramePack/discussions/459

  • When you first time try to load F1, it will download models into hf_download sub folder inside FramePack

    • I kept all of the existing features even timestamp based prompt working

    • I did few tests and all working but if there are errors please let me know

  • Now batch processing will sort by name ascending and process

    • It has a new library so download latest zip file and run installer for update

  • Used Negative prompts will be saved in metadata txt files now

26 April 2025 Updates V44

  • With v43 Teacache was disabled with an error

    • Please run updater from V43 to get updated, just run installer

  • 2 new bat files added to zip file

    • Uninstall_Sage_Attention_For_Better_Quality_Slower.bat

      • Sage Attention is fastest but reduced quality a certain degree

    • Uninstall_Sage_and_Flash_Attention_For_Best_Quality_Slowest.bat

      • Flash Attention is faster but reduces quality slightly

    • xFormers is slowest among above but quality best

      • Also xFormers supports RTX 2000 series and below

26 April 2025 Updates V43

  • Installer file updated and now includes official Torch 2.7 - not anymore nightly version

    • I also have compiled xFormers and Flash Attention for this version and they will be used from now on

    • You need to delete venv folder and run installer so it can update all libraries

  • I have completey remade LoRA loading and deloading logic and now it works without VRAM increase

    • LoRA dynamic loading and reloading was breaking Preserved Memory system which was causing model to use full VRAM

  • Download latest zip file and run installer. Before doing this you can delete venv folder for latest libraries to be installed

  • Batch processing Gradio error fixed

  • If you open other apps etc during the generation, your FramePack will start using shared VRAM and it will become very very slow

  • Preserved Memory description updated as below

    • GPU Inference Preserved Memory (GB) (larger means slower on big VRAM but you need to have it large enough to not use shared VRAM (on low VRAM machines) or it will be extremely slower)

      • Set this number to a larger value if you encounter OOM. Larger value causes slower speed. Do not set this very low that will cause shared VRAM usage

  • LoRA convert bat file removed since it is not needed anymore as it is done dynamically now

24 April 2025 Updates V40

  • For older GPUs (RTX 2000 and below) an installer fix file added

    • After installing run Older_GPUs_2000_And_Below_Install_Fix.bat if you have RTX 2000 or below GPUs

  • High VRAM mode disabled (was only activated if GPU was more than 60 GB VRAM)

    • The reason is, there was a bug which made it ultra slow

  • GPU Inference Preserved Memory parameter was not being used at VAE decoding and it is fixed

  • The app processes videos as iterations and the iteration count and how many frames processed at each iteration is directly related to Latent Window Size

  • Latent Window Size : Controls generation chunks. Affects section count and duration (see info above prompt).

  • Default is 9 but you can change it. If you increase it, it will do lesser iterations to generate videos, thus shorter time

  • If you reduce it, it will do more iterations, thus more generation time

  • I didn't thoroughly tested yet the impact

  • We also have added Set Window for ~1s Sections button

  • With this button It will try to auto set Latent Window Size so that each iteration will be almost equal to 1 second

  • With this way your timestamp prompts will work more accurately hopefully

  • Example prompt for 30 second video is like below

    • [0] a man talking

    • [5] a man crying

    • [10] a man smiling

    • [15] a man frowning

    • [20] a man sleepy

    • [25] a man going crazy

    • How they are processed as below

      • [0] a man talking : between 0-5 seconds

      • [5] a man crying : between 5-10 seconds

      • [10] a man smiling : between 10-15 seconds

      • [15] a man frowning : between 15-20 seconds

      • [20] a man sleepy : between 20-25 seconds

      • [25] a man going crazy : between 25-30 seconds

  • Timestamp processing improved

  • This new timestamp prompts will only work when multi-line prompt is not active

  • This should work both for single image processing and also batch folder processing

    • So in your batch folder prompt txt files you can write such prompts

  • Now you can set GPU Inference Preserved Memory = 0

    • Useful for big VRAM GPUs and when doing inference on lower resolution

  • Please download latest zip file and run installer bat file. It should be pretty quick on existing installations

23 April 2025 Updates V31

  • The reported save error bugs fixed

    • Now both save last frame and RIFE should work please report any errors

  • I also saw some Chinese guy distributing my app, I have contacted him I hope he stops this or accept my terms

23 April 2025 Updates V30

  • CFG scale and negative prompt features added

    • You need more than 1 CFG for negative prompt to be enabled

    • CFG means how much it will follow your prompt

    • There is also CFG Re-Scale which i don't know what kind of effect it makes

    • Default CFG Re-Scale = 0 and CFG = 1

    • Enabling negative prompt may increase generation time with CFG > 1

  • Presets feature implemented

    • Please test it thoroughly and let me know if any errors

    • You can save your presets and load them, it should save and load everything except main prompt

  • Practical RIFE 2x and 4x frame interpolation implemented

    • This requires to download v28 zip file and run installer again

    • FPS increased videos will be saved with name like this 2025_04_23_02_56_26_130_extra_FPS.mp4

    • It should work for both single image and batch image processing

  • Save last frame feature added after requested

  • Moreover, FPS directly changes how many frames processed

    • So if you set 60 FPS, it will generate double amount of frames thus double time

    • If you set as 15, it will generate 15 frames every second

    • Hopefully when I add RIFE 2x and 4x FPS this can be leveraged, speed up generation twice with 15 FPS

21 April 2025 Updates V25

  • Save Individual Frames feature implemented

  • They will be saved inside outputs\individual_frames folder and should work with both batch and single image processing

    • At batch processing they will be saved inside user given folder/individual_frames

  • When saving files new file saving format is like 2025_05_07_12_42_12_333

    • We have removed seed from names

  • Now multi-line prompting works with batch processing as well

    • Make prompts txt files inside your batch folder and they will be auto looped - each line is a new prompt new video

  • Teacache slider added : default is 0.15

    • You can increase or decrease

    • Set 0 to disable it

  • Just run Windows_Update.bat and it should work

21 April 2025 Updates V23

  • Multi-line prompting added

  • When this is enabled, it will generate each line as a separate new video

  • For RunPod and Massed Compute same just run install command again (use latest zip file)

20 April 2025 Updates V21

  • We have implemented LoRA support

  • Currently diffuser type LoRAs are supported but you can try Windows_Convert_Loras.bat and convert them

  • Put LoRAs into loras folder and hit refresh

  • We have implemented beginning and ending frame feature

  • It is working amazing and so easy to use

  • Beginning and ending frame feature supported at batch processing as well

    • The logic is that have same file name e.g.:

    • Beginning folder 1.png, ending folder 1.png

  • We have new requirement so run Windows_Install.bat for update, it will update don't worry not reinstall from scratch

20 April 2025 Updates V19

  • Now you can even set output FPS of videos

  • This will not change your generation speeds since model is built for 30 frames for every second but it will change output video FPS

  • New resolutions implemented 1200, 1320, 1440

  • Now generation times will be logged into save prompt txt files. e.g.

    • Generation Time: 1 min 38 seconds

    • Total Seconds: 98 seconds

  • 1440 bucket has 1920x1088 and 960 bucket has 1280, 704

20 April 2025 Updates V16

  • New resolution buckets added

  • Lower resolutions are way faster and still working great

  • It will also display which resolution bucket your image fits into according to the resolution you pick

  • Available resolutions 640 (like 640x640 - this is default original res), 1080, 960, 840, 720, 480, 320, 240

    • When you select 640 or below resolution the GPU Inference Preserved Memory will be auto set to 6 GB

      • For lower than 640 res you can try lower values for faster inference like 5 GB , 4 GB, now as low as 2 GB possible

    • For 720 = 10 GB , For 840 = 12 GB , For 960 = 14 GB , For 1080 = 16 GB

      • If you see it is using too much VRAM increase above amounts like with 2 GB more

      • Use nvitop and make sure at least 1 GB of VRAM is free

  • If you get out of VRAM error or if it uses shared VRAM (control your GPU watt usage), increase Inference Preserved Memory

  • Square resolutions added like 640x640, 480x480

  • Start generation and End Generation buttons toggle on and toggle off issue fixed

  • Displaying intermediate videos issue fixed

  • Now you will see preview videos during generation

  • During batch processing no preview or final videos were displayed and this issue also fixed

  • Now you should see both intermediate videos and final videos in batch processing

  • Just run Windows_Update.bat and it should work

19 April 2025 Updates V12

  • Batch processing arrived

    • Number of images now working with batch processing

  • Just run Windows_Update.bat file

  • The logic so simple, give your input folder, output folder, and thats it

  • It will read all the images inside given intput folder, also if you want custom prompt for each image logic is like below

  • image_file_name.txt. The image extension doesnt matter. It will read it

  • Moreover a new feature implemented : Save Processing Metadata

  • This will save the following metadata as below

    • There were few errors with saving now fixed with v12

Prompt: a warrior vs a dragon

Seed: 29075816

TeaCache: Enabled

Video Length (seconds): 5

Steps: 25

Distilled CFG Scale: 10

17 April 2025 Updates V9

Please upvote and leave a comment if possible : https://www.reddit.com/r/StableDiffusion/comments/1k1m4j6/15_wild_examples_of_framepack_from_lllyasviel/

Tutorial video : https://youtu.be/HwMngohRmHg

  • Output folders greatly organized as below

  • Number of generations implemented - lets say you generate 3, it will generate 3 videos with same settings with random seeds

  • Random seed implemented

  • If you use static seed, number of generations will increase seed by 1 and generate

  • Save as Gif, APNG and WebP added

    • Be careful generations taking a lot of space

  • Interface improved

  • Statistics improved, now will display how much time passed and estimated time to complete

How To Install and Use:

  • Total installation is around 50 GB

  • Works as low as 6 GB GPUs

Windows:

  • First watch above requirements tutorial and follow every step

    • Extract zip archive into a drive like c:/FramePack_v1

    • Longer directory path or space character or special character in folder paths causes issues

    • Double click and install with Windows_Install.bat

    • Then run with Windows_Start_App.bat

      • If you get model download error open Windows_Start_APP.bat with a notepad editor and make "HF_HUB_ENABLE_HF_TRANSFER=0"

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

Damjan Žakelj

I get few errors when installing in RunPod, fist one is in the star of install: Cloning into 'FramePack'... fatal: unable to access 'https://github.com/FurkanGozukara/FramePack/': Failed to connect to github.com port 443 after 133653 ms: Connection timed out ./RunPod_Install.sh: line 10: cd: FramePack: No such file or directory, And when i try to start it, I get: bash: cd: /workspace/FramePack: No such file or directory python: can't open file '/workspace/app.py': [Errno 2] No such file or directory (venv) root@d0968809e9a8:/workspace#

cool1

It's still downloading currently, but it looks good if it can generate up to 2 minute videos and doesn't need much RAM. Though I don't know about the speed. But unless there's Lora capability won't it likely get inconsistent results after a short time (unless a person/character etc. is always facing in roughly the same direction)? Is there not a Lora option?

SECourses

no lora yet. it is really great try it. i am adding new features already

SECourses

your pod simply failed. Failed to connect means pod internet error . get a new pod .

SECourses

you are welcome. i installed on runpod twice works perfect :) unless your pod broken haha

Hockey

This is better than the Wan 2.1 software. =)

SECourses

this is really great. adding more features already :D

DanO..

Better than Wan 2.1? That's a big claim. I'm getting some great stuff with loras. I must check it out then.

Hockey

I can only speak from my own experience, I tried to animate an image and it was just messy, tried several times. Now it worked immediately. Lora support would be awesome.

SECourses

i cant say better but faster and easier :D

SECourses

i agree it depends on case and loras would be great

Jim Coupe

Watched the vid, installed the app, works great on my 5090, thanks a lot. One small suggestion: would it be possible to add total generation time so as we play with different settings we can get a sense of time vs output quality?

SECourses

good idea working on it right now for you

SECourses

just added some statistics can you update to v9 and try and let me know

AI Tools Forge

This is the best video generator I tried till now! Thank you!

SECourses

thank you so much. hopefully i will add more features.

Jim Coupe

works great, thank you very much. I like how it shows you the ETA for the job as it goes, and the total generation (and in total seconds and mm:ss) Now I can play with the settings and find the best quality vs time for me and easily measure the difference. Thank you. example of output for the curious: Progress: 88% | Sampling 22/25 (Gen 1/1, Seed 1455904109) | Elapsed: 3.2 min | ETA: 25.9 sec Progress: 92% | Sampling 23/25 (Gen 1/1, Seed 1455904109) | Elapsed: 3.2 min | ETA: 24.8 sec Progress: 96% | Sampling 24/25 (Gen 1/1, Seed 1455904109) | Elapsed: 3.2 min | ETA: 23.5 sec Progress: 100% | Sampling 25/25 (Gen 1/1, Seed 1455904109) | Elapsed: 3.3 min | ETA: 22.3 sec 100%|████████████████████████████████████████████████████████| 25/25 [00:31<00:00, 1.24s/it] Section completed in 31.15 seconds Offloading DynamicSwap_HunyuanVideoTransformer3DModelPacked from cuda:0 to preserve memory: 8 GB Loaded AutoencoderKLHunyuanVideo to cuda:0 as complete. VAE decoding completed in 16.09 seconds (0.435 sec/frame) Unloaded AutoencoderKLHunyuanVideo as complete. Saved MP4 video to D:\AI\Video\FramePack\FramePack\outputs\250417_232828_147_1038_seed1455904109_37.mp4 Saving operations completed in 1.55 seconds Decoded. Current latent shape torch.Size([1, 16, 37, 104, 60]); pixel shape torch.Size([1, 3, 145, 832, 480]) Generation 1/1 completed in 215.25 seconds

Ammar Salem

Can this be used through API to use through http request ? As Always, thank you so much for all the hard work providing us the latest and greatest one lick apps 🙏

SECourses

Gradio has api itself so check bottom of your gradio app you will see Use via API

SECourses

currently you can open your 127.0.0.1 url on different tabs and gradio will automatically queue your every generation every setup. hopefully i will add more features and batch processing of folder where you will be able to set different prompt for each generation

Paul Knaggs

Really good, worked like a charm on my RTX 5090

Paul Knaggs

I have a new 50 series GPU. I want to make a Lora how and what platform would you recommend?

DanO..

I'm enjoying experimenting with this. I can't wait for loras and possibly WAN support because it's clear to me WAN is superior. I'm getting a second or two less than what I ask for a couple times. Just a heads up. I got a 9 second video when asking for 10 and a 4 second when asking for 5.

Jesen

Interesting. After generating one second of video, I get: Offloading DynamicSwap_HunyuanVideoTransformer3DModelPacked from cuda:0 to preserve memory: 8 GB Press Any Key to continue.... and it errors out in the interface. Is this a settings issue? 4060 16gb

SECourses

hello. you are out of RAM. please set virtual RAM to like 50 GB on your fast disk and it should get fixed

SECourses

ye the duration not exactly precise i noticed that :D i hope loras will arrive soon and i will add more features like batch folder processing

SECourses

for lora making it is in my research not ready yet :(

SECourses

awesome. i also have rtx 5090 to do actual testing and development. i have rtx 3090 as well to test on older gpus :D

SECourses

it can be related to ffmpeg. which ffmpeg you are using? did you follow requirements video ffmpeg?

Jerad Benge

This is a lot of fun to play with! Thank you for the installer.

SECourses

you are welcome. just implemented batch processing as well with custom prompt support :)

SECourses

batch added please try v10 and let me know

cool1

I also noticed the frame duration not being the same as the one asked for in the GUI (ie. asking for 5 second video got a 4 point something second one I think) which I posted about I think. If Grok is asked about it it creates a long reply about it with suggestions and the conclusion was: "The incorrect video duration...is likely due to rounding errors in total_latent_sections, frame subtraction in num_frames, and potential frame loss during overlap blending. The suggested fixes—using ceil for section calculation, removing the -3 frame subtraction, trimming history_pixels, and verifying encoding—should ensure the output video closely matches the requested duration...". Though I don't know if it's suggestions would work. Hopefully we will also be able to specify things like resolution x & y too. Though I don't know if the resolution is limited to quite low because it was only trained on lowish resolution videos. If so, maybe when WAN 2.1 is used in a FramePack system we can have the higher resolution (even though it would be lower speed at higher res, the VRAM needed wouldn't increase with video length like normal WAN 2.1 video generation. On one page about the FramePack app it says Teacache can lower quality (or sometimes and showed an example where it lowered it). And it said it could be disabled for best quality (though that will likely lead to a lot longer render times). Though another video app I think asked for a value for TeaCache. So maybe if it does lead to lower quality, if we had a setting for TeaCache value then that could be changed/lowered so it didn't reduce quality as much (for videos where it reduced it) but you could still have it enabled to make it at least faster than rendering without it.

Draxel

This is neat. Anyway to make the outputs less "glossy?"

SECourses

thanks. i dont know will check out

mastermk

Hi Furkan. I saw that it doesn't update the Finished Frames, it only shows the first generation and the End Generation button is disabled.

DanO..

I think I may be talking about the same thing as an earlier patron. After updating the film strip preview goes away after the first second is generated and it seems to just show one second finished at a time instead of the cumulative video generation. It also looks like it has completed while it's still generating video.

SECourses

with v13 fixed thanks for reporting

SECourses

with v13 fixed thanks for reporting

mastermk

Thanks Furkan

Han Vu

Hello Furkan, recently the new LTXV models are insane. They are producing videos in 60 seconds. One extremely important feature of the workflow, is that it connects to an LLM such as chatGPT or Gemini, in order to make dynamic videos on the fly each time. Can you please implement something like this? Like an option to send a prompt to LLM (with image), it will analyze and return with a prompt for the image? This will require a field for an API key as well. This is so important for dynamic video generations. It is already working so well in the LTXV workflow, would be great to see it implemented for WAN or Framepack, which it creates varied videos of the same image.

Han Vu

This is the example prompt being used, it's amazing when it returns a video prompt: Ignore all prior context and outputs. You'll start fresh. You are an expert cinematic director and prompt engineer specializing in text-to-video generation. You receive an image and/or visual descriptions and expand them into vivid cinematic prompts. Your task is to imagine and describe a natural visual action or camera movement that could realistically unfold from the still moment, as if capturing the next 5 seconds of a scene. Focus exclusively on visual storytelling—do not include sound, music, inner thoughts, or dialogue. Infer a logical and expressive action or gesture based on the visual pose, gaze, posture, hand positioning, and facial expression of characters. For instance: - If a subject's hands are near their face, imagine them removing or revealing something - If two people are close and facing each other, imagine a gesture of connection like touching, smiling, or leaning in - If a character looks focused or searching, imagine a glance upward, a head turn, or them interacting with an object just out of frame - If it is a far/distant shot of a subject(s) from far away, then the description should describe how these two subjects are interacting (talking, playing, petting, etc that is reasonable for the image) Describe these inferred movements and camera behavior with precision and clarity, as a cinematographer would. Always write in a single cinematic paragraph. Be as descriptive as possible, focusing on details of the subject's appearance and intricate details on the scene or setting. Follow this structure: - Start with the first clear motion, still shot or slight camera cues (pan, tilt, rotate, zoom in/out, slow motion, hand cam, drone cam), be sure to randomly choose between these to vary between the shots. - Beginning emote: If the subject is a person or animal, include an expression of emotion, choose randomly one emotion, (e.g. such as but not limited to, happy, laughing, smiling, talking, staring off into distance, looking away, looking up, looking down, staring, serious look, chuckling, slight smile, sad, angry, curious, suspicious, confused, peaceful, at peace, closes eyes, friendly gesture, friendly look, calm) - Build with gestures, body language, expressions, and any physical interaction - Detail environment, framing, and ambiance (be sure to include slight breeze or gentle wind if environment has hair, grass, cloth) - Ending emote: If the subject is a person or animal, finish off with transition expression of emotion, choose randomly one emotion, (e.g. such as but not limited to, happy, laughing, smiling, talking, staring off into distance, looking away, looking up, looking down, staring, serious look, chuckling, slight smile, sad, angry, curious, suspicious, confused, peaceful, at peace, closes eyes, friendly gesture, friendly look, calm) - Finish with cinematic references like: “In the style of an award-winning indie drama” or “Shot on Arri Alexa, printed on Kodak 2383 film print” If any additional user instructions are added after this sentence, use them as reference for your prompt. Otherwise, focus only on the input image analysis:

Han Vu

Something like this is produced but its totally different everytime: The camera slowly pans up from the dark fur of the resting dog to its eyes, which gleam with curiosity in the low golden light of the setting sun. As the soft breeze ruffles the grass around, the dog's ears tilt forward with a hint of intrigue, capturing interest on something just beyond the frame. Its tail gives a subtle wag, disturbing the amber blades that cradle its form. The shot then gently zooms out, revealing distant mountain peaks silhouetted against the glowing horizon. The dog’s expression transitions to serene attentiveness, focusing intently as if awaiting an unseen cue. Shot on Arri Alexa, printed on Kodak 2383 film print, in the style of a poignant nature documentary.

Han Vu

This results in multiple shots of the same image, producing varied videos that can be used to splice or edit together for longer length videos.

Han Vu

Additionally, please implement the LTXV (new 0.9.6 and distilled) work flow. It's really amazing. Amazing videos are produced in 60 seconds on RTX 3090!

stphnvdb321

Another great app! Thank you. I just updated to v16. If I chose 720p, the output video is only 928x544p (my input picture was 1280x720p). A bug?

cool1

It could be done with local free LLMs rather than closed source/paid non-local ones (like Gemini/ChatGPT). eg. CogVLM2 image captioner (that a gradio for was posted here) can be asked to create a video description based on an image you give it. So something like that could be integrated into a video generation app (such as FramePack). And if it isn't integrated, you could still use CogVLM2 in batch mode to create video captions for each source image, though it would only create 1 video caption per image - so you couldn't create multiple different video ideas from one image as easily (unless you created multiple folders of the same images and re-captioned).

Han Vu

Running locally only makes sense if you plan to produce the prompts in advanced. You want to leverage cheap or free Gemini (on the fly) so that you dont eat up vram or load and purge vram, over and over, it's slower. Just use free Gemini and its done instantly.

mastermk

Hi Furkan. Is it possible to insert the negative prompt? Then another question, How many tokens can the positive prompt handle?

SECourses

no negative prompts and it can handle 128 token (not 128 character)

SECourses

this 720p is 720x720p this is how Hunyuan resolutions are mentioned :D

SECourses

certainly api based llm can be implemented to get prompt. which api you have openai?

SECourses

how Gemini is free can you give more info? we could add it

stphnvdb321

I never used Hunyuan before. It would be awesome, if FramePack could use Wan instead, but I think this can't be done. :(

SECourses

i am following and i think we might get wan 2.1 as well in this app

mastermk

Sorry for the many questions, During the generation of 5-minute videos, in the first generation the scene starts with a certain speed and dynamism. In the remaining generations, the actions tend to become static. In the generations of 10-minute videos, towards the end, the movements are almost zero, similar to the initial image.

SECourses

yes i noticed this too. i wonder if can be solved, looking for a solution

cool1

Thanks for the resolution adding and batch mode adding. Maybe an option to specify the compression amount could be added too. I asked it for a 720p resolution video and gave it a 1024x1024 start image and it's creating intermediate vids of 704x704 pixels. Shouldn't it be 720x720 for the intermediate vids? The end video it created was 704x704 pixels to (when asked for 720p).

cool1

You woudln't need to generate the prompts in advance if a call to a local LLM (such as CogVLM2 any other multi modal one that could do it, including any one that uses lower VRAM) is integrated into the video app (eg. FramePack). then you don't need to generate all the prompts in advance. You can ask the video app to analyse the image generate the prompt for the specified length of video. It could generate multiple prompts for it so different videos could be created from the same image. All local and all free and no connection to paid AI models needed and no need to sign up on any site. I didn't think Gemini is free and ChatGPT needs sign up etc. There should be an all open source free option. Also For OpenAI API usage that's not free according to perplexity.ai, you need to pay for credits for it (unlike a free local one).

SECourses

no it has to be divisible to 32. we already have quality Video Quality High: Best quality, Medium: Balanced, Low: Smallest file size, Web Compatible: Best browser compatibility I may add slider but that makes people confused i dont know should i add or not.

Bjoern

Thanks for the installer. Any way that you can make a Kaggle notebook for it?

cool1

Thanks. I didn't see the best/medium/low. etc. setting. Adding a quality slider could allow it to be more precise but it's okay if not. If it needs to be divisible by 32 coudln't that be taken account in the selection dropdown box. It allows us to select a resolution of "720" but then creates a video (when given a 1024x1024 pixel image) of 704x704. couldn't it put only the ones it can create in the dropdown selection list (eg. 704 instead of 720 if 720 can't be output, 1088 instead of 1080 if 1080 can't be output)? 840 is in the list too and that isn't divisible by 32 either so maybe that could be replaced by one that is divisible by 32 in the list.

SECourses

well i am not sure the T4 doesnt support modern attentions that is the biggest problem. we are using modern attentions here. i need to try

SECourses

this is how it works. also 720 means 720x720 pixels. you can use higher res too

cool1

But 720 isn't divisible by 32. That's probably why it gave me 704 instead. And when I gave it a 1280x720 source image and tried different resolutions, I still couldn't get it to give me approx the same resolution, without cropping. When I gave it a 1280x720 source image and selected "960" resolution it gave me a 1280x704 pixel video (so the height is cropped). The only next size up in the list is "1080" (which also isn't divisible by 32 so I assume it will convert to 1088). But that might take a long time rendering. If it has to be that each axis needs to be divisible by 32, then, for when we give it 1280x720 pixel source images, and that wouldn't be 32 pixel divisible on each side, could there be a 1312x736 pixel option added (for when we give it 1280x720 source images and want about 1.777:1 aspect ratio)? 1312x736 is divisible by 32 on both axis and is about 1.7826:1 aspect ratio, so quite close to the original so I assume it could work okay. edit: selecting the next size up of "1080" gave me 1376x832. I assumed it would give me 1920x1088 as that's the closest to 1920x1080 while being divisible by 32 on each axis. So maybe that could be used (but that's about 1.654:1 and not as close to 1.777:1 as 1312x736 is, so 1312x736 might be a better resolution for 1280x720 source images if that was supported as it would maintain a closer aspect ratio and render faster) but when I saw it could go to 1080 I assumed it would be 1080p as that's what it said somewhere on reddit. Or maybe it limits it because the source image is a lot lower than 1080p.

SECourses

you are welcome and new features coming. like lora and ending frame :D

Han Vu

You missed the whole point that it takes up vram, do you really want to add another step that eats up vram? Gemini has daily free tokens that make it virtually free. Just sign up at AISTUDIO and get an API key. It's been virtually free for 2 years since it came out. The daily limits are ridiculously high that I've never maxed it out, even doing insane requests.

Han Vu

I can do it all locally as well, but it takes time integrate it separate from Furkan's program, it costs vram to run it first to get the prompt, you're asking too much for other people that may have less than 24gb vram, would just be unnecessary. I guess if you want, you can integrate deepseek locally to satisfy this completely local solution, but it simply adds more step and barrier to integrate and another factor that could fail if the users computer isnt optimal or the optimal model isnt chosen (which adds another point that these models add additional download space), because an API is practically instant. API solves this for 99% other people. Yes I have ollama, LMStudio, I still wouldn't want to use it for this when Gemini is free and instant.

Han Vu

Google AI Studio https://aistudio.google.com Google AI Studio is free to use, virtually all of their models are free to use, with the even the latest models! They're called experimental, but they work like 99% like production models. They have daily limits, but the daily limits are virtually absurdly high, even when using their latest Pro models. Youll need to use their vision based models like Gemini 2.5 Flash

cool1

Sometimes in places in a rendered 3d cgi cartoon style video I see some vertical banding on the 3d cartoon man's greyish chin (that wasn't in the original source image). This is when rendered at 640x640 or 704x704, both at the "high" (highest) quality setting. There's also some artefacts on the hand at one point. I don't know if it's to do with the video compression or whether it's because it's 8 bit or how it's processed internally (eg. whether that always has enough bit depth). Though maybe it could have a slider for video quality where the highest setting was higher than the current "high" value. Just in case it's that. Maybe it could be related to TeaCache it that's lower quality sometimes but I haven't tried disabling that yet as it would probably take much longer to render. If it is that then a slider/value for TeaCache where we could reduce the effect of it could be useful.

SECourses

it is related to model. config is really high quality if video_quality == 'high': options = { 'crf': '8', # Lower CRF = higher quality (18 is high quality) 'preset': 'slow', # Slower preset = better compression 'pix_fmt': 'yuv420p', # Most compatible pixel format 'movflags': '+faststart' # Allows video to start playing before it's fully downloaded }

cool1

Thanks. Yes you're probably right that it's the model. Though a setting slightly less compressed (eg. maybe an option for crf 7 or 6 could be added with a slider) might help too, especially if it gets re-compressed when it keeps combining the intermediate videos. Maybe the YUV 4:2:0 might have an effect but it's mostly grey there in the chin where it shows the banding so it probably isn't a colour sampling issue there. Maybe higher res might help and then downscaling or maybe there's a way of debanding with some filter. Or maybe there's nothing much that can be done (eg. until we have models that don't do that, eg. that use higher bit depths - but we probably would need higher VRAM GPUs for that).

SECourses

crf 8 is actually already overkill :D below 12 is pseudo quality

cool1

Thanks for the update, with the Lora support and Start and End frame settings. For the Lora, what do they need to be trained on? Trained on Hunyuan AI video or FramePack? Is there a way we can train them or would we need to use existing Lora's someone's trained on this video model?

SECourses

i tested few hunyuan and they didnt do a difference but try and see :D

cool1

Thanks. I'll check. Ideally we could also train our own to use too it there's a way for that that works on consumer GPUs.

DanO..

So, we need to create the "loras" folder, yes? And what is "Windows_Convert_Loras.bat"? Where do we get it? How do we know what type of loras are on CivitAI? And do the regular CivitAI loras work with i2v? Thanks for all the hard work!

SECourses

we dont know and loras not working properly atm. who claims they implemented loras working i am testing all but not very good :D wait me to make new announcement and just try atm and see if works or not :D

DanO..

Yeah. I don't know if it's connected but my machine maxed out all resources and sort of froze so I had to power cycle. So... didn't work for me! Waiting for an update from you.

David Mudd

It would be great if we could access the individual frames (either as png or jpg) that were produced before they were converted to a video. Is that possible?

SECourses

definitely possible let me add it

100Sachen80

Hi Furkan, can you add feature from WAN "Practical-RIFE" frame multiplier?

stphnvdb321

I have a request: Can we have a slider for TeaCache like in your Wan app to adjust the strength? To my opinion it is set too high by default and fingers really get distorted

Stevie Tee

Thank you for the great updates ! Just one thing I've noticed though, has the original developer got the "preserved memory" settings the correct way around ? Because it seems that when I increase the setting to 12 (I have a 12GB 3080 TI card) that the speed increases not decreases, is "larger slower" ?

SECourses

that means your system using too much vram therefore you are using shared vram. perhaps i can increase default to 8 from 6

Stevie Tee

I'm a bit confused. Before using your improved version, I was using a FramePack version in Pinokio, the default on that one was 6. I do have 256 GB of system RAM (not GPU RAM) (using a HP Z series workstation) so perhaps the system VRAM is doing the heavy lifting. When I lowered the preserved memory setting to 2, to test, it slowed down much more, that's why I thought perhaps the developer had set it the wrong way, or gave incorrect instructions in the preserved memory dialogue box.

SECourses

actually the description is accurate GPU Inference Preserved Memory (GB) (larger means slower) larger means slower. when you make it 2 GB, you are actually using shared vram since your VRAM is not being sufficient. this larger would make it slow like in 80 GB machine where shared VRAM would not be used at all :D so until you hit your shared VRAM, lower is faster, but once you hit shared VRAM, lower will be slower

SECourses

i just added it with last update. please update to see and test.

Stevie Tee

Ha ha ! Thank you, I think I understand now.

cool1

In version 21 of the app (and I haven't seen it say it's changed in more recent ones), sometimes the generation is going, but the "Start Generation" is enabled and the "End Generation" button is disabled, so you can't stop it. Maybe it happened because of starting and then stopping a few times. Though I just found out you can still click "stat generation" (when it's already started) and then the "End generation" button becomes enabled and you can click it and stop the generation that way.

Jim Coupe

This is a good add. A small tweak would be the option to save the last frame of the video for future continuation if needed.

SECourses

update to latest version we improved them

SECourses

just implemented update to use :D

Leonid Wolf

The latest version is very working. Now I can adjust the video resolution with my GPU 12gb memory for productive generation, Thank you for the build.

SECourses

just added that option for you :D update to v26

Samuel

What about getting prompt metadata from the images ¿

SECourses

we dont have it. currently you can read from txt files which we save. metada data image thing is adding too much work with too little benefit. i plan to add more impactful features tomorrow hopefully

David Mudd

Just tested! Works great! Thank you!

Ken

This turning out to be wonderful. Will you be taking on SKyreels 2

TaTaKae

I know LoRA is still experimental, but I’ve gotten some good results using a few LoRAs on RunPod—even though they currently use a crazy amount of VRAM. Would it be possible to add support for multiple LoRAs and the ability to adjust their individual scale? If not, I totally understand. Thanks for the amazing work!

Art of 3D Rendering

Can you add the modification you post in reddit group to choose the resolution size for the video output?

SECourses

we already have resolutions. or you mean something else? we have "1440","1320","1200","1080","960","840","720", "640", "480", "320", "240"

SECourses

Can you tell me which LoRA worked? I will try to fix that vram issue

SECourses

v25 zip file definitely has it. can you reinstall and show me your started screenshot? just run the installer - you can also send me install logs monstermmorpg@gmail.com

Ec Jep

Thx for this amazing work! In an earlier version you could adjust the steps during vid generation - now they are gone. I am trying to use start and end frame and the generation use one frame for the start and then starts backwards from the end frame. is there a way to have it focus on the start frame as well so its it more balanced on start and end frames?

SECourses

thanks. the steps count still there look interface carefully. i am working to add timed prompt so for each second you can set a different prompt

Deniz Nasif

is there a simple way I could just update to v25 if I already have v23 installed without having to install v25 from scratch?

SECourses

this zip file updates to latest . currently latest version is v23 but zip file is v25 :D so you are up to date. sorry for confusion

Deniz Nasif

ok thanks, i can't see any of the latest features though, such as teacache slider or save individual frames

SECourses

extract zip file into same folder, overwrite and run installer and let me know logs

SECourses

ok you need to see v26 on gradio interface not v23. please do what i said and report me back

Deniz Nasif

looks like there is a hash mismatch when running the installer in the latest zip file ERROR: THESE PACKAGES DO NOT MATCH THE HASHES FROM THE REQUIREMENTS FILE. If you have updated the package versions, please update the hashes. Otherwise, examine the package contents carefully; someone may have tampered with them. torchvision from https://download.pytorch.org/whl/nightly/cu128/torchvision-0.22.0.dev20250422%2Bcu128-cp310-cp310-win_amd64.whl: Expected sha256 8ee0514d2cddf2b615ce0a85415d9c98f99e98cbd185c22d2a23e2d849dce06e Got 0bf6ad5fc142196eb0559a44a47b9787e97a7777df9140f74474f262aa17fe9f

SECourses

it is related to your pip. open cmd and do pip cache purge if still not fixed turn off vpn, antivirus etc. if still not fixed install and use while WARP vpn is open. one of these will likely to fix

Ec Jep

I am using v25 and still don't see the interface.

SECourses

please get v27 zip and run installer. if you still get error make a fresh install and move hf_download and models folders

Deniz Nasif

Your latest installer version you just uploaded (v26_V2) worked fine. I was able to move over the models from V23 to save me having to download again after I had finished installing V26.

umut yilmaz

Getting better and better everyday <3 Maybe we can add and sync voice & sound one day.

SECourses

yep it is on my research too but probably another app. currently adding RIFE 2x and 4x FPS

SECourses

changed remote repo to bitbucket atm. i hope it gets reinstated asap

Proyecto Hermenautas - navegar la astrología

Hi, I'm new to your Patreon and wanted to congratulate you on your excellent work. I hope the problem with your GitHub account is resolved soon. On the other hand, I just downloaded and tested the latest version. Overall, it works great, the extra control of indicating actions using timestamps is great! But I found some problems: 1 - Unlike previous versions, it no longer previews the video while it is being generated. 2.- It does not generate the .txt files with metadata. 3.- Despite activating the RIFE option, it does not generate any extra_FPS.mp4 file. Greetings and again congratulations on your excellent work!

Jim Coupe

I had this issue when I had 'Save Last Frame of Generation' checked. Unchecking that allowed the txt and RIFE to save correctly. I reported this in the discord I am sure it will be fixed soon.

Proyecto Hermenautas - navegar la astrología

Checking the terminal output, I see the following warning appears: "No output file was generated due to an error" and the video it generates is one second shorter than the total generation time I set. I hope this helps to find the problem.

Rudbby

On Runpod I am getting 504 Gateway Time-Out for Gradio on the new version

andersefp

Outstanding work, very well done!!! There seems to be a problem with using the "different prompt for every second" if you generate videos longer than 5 seconds... It seems to take last prompt within 5 seconds and do the whole generation with that prompt. Or am I missing a step?

SECourses

well if you generate 10 seconds and your last is after 5 seconds, the last 5 second should be that one. you are asking that right?

SECourses

related to runpod pod. they tends to be broken a lot. please get a new pod

SECourses

working on last frame bug . also sometimes preview may not be displayed it is due to gradio nature not excellent.

SECourses

thanks for reporting working on it

SECourses

actually not shorter. if you look at video you will see like 4.8 seconds or 4.9. but not exactly 5 seconds true sometimes

andersefp

Correction: If a timestep is set to occur after 5 sec, it seems to get ignored.. So if I generate a 10sec video with 0] a man is crying [4] a man is smiling[7] a man is screaming- the "a man is screaming" gets ignored.

Damien Jourdan

an easy suggestion maybe : a button to swap star and end frame

Weronika

i have problem WARNING. For this auto installer to work you need to have installed Python 3.10.x, Git, FFmpeg, cuDNN 8.9 and 9.4+, CUDA 12.4+ and C++ tools This tutorial shows all step by step : https://youtu.be/-NjNy7afOQ0 'git' is not recognized as an internal or external command, operable program or batch file. 'git' is not recognized as an internal or external command, operable program or batch file. 'git' is not recognized as an internal or external command, operable program or batch file. Python launcher is available. Generating Python 3.10 VENV Python 3.10 not found! Installed Pythons found by py Launcher for Windows -3.9-64 * Requested Python version (3.10) not installed, use -0 for available pythons System nie może odnaleźć określonej ścieżki. Requirement already satisfied: pip in c:\users\user\appdata\local\programs\python\python39\lib\site-packages (25.0.1) Looking in indexes: https://download.pytorch.org/whl/nightly/cu128 Requirement already satisfied: torch==2.7.0.dev20250311 in c:\users\user\appdata\local\programs\python\python39\lib\site-packages (2.7.0.dev20250311+cu128) Requirement already satisfied: torchvision==0.22.0.dev20250312+cu128 in c:\users\user\appdata\local\programs\python\python39\lib\site-packages (0.22.0.dev20250312+cu128) Requirement already satisfied: torchaudio in c:\users\user\appdata\local\programs\python\python39\lib\site-packages (2.6.0.dev20250312+cu128) Requirement already satisfied: filelock in c:\users\user\appdata\local\programs\python\python39\lib\site-packages (from torch==2.7.0.dev20250311) (3.16.1) Requirement already satisfied: typing-extensions>=4.10.0 in c:\users\user\appdata\local\programs\python\python39\lib\site-packages (from torch==2.7.0.dev20250311) (4.13.0) Requirement already satisfied: sympy>=1.13.3 in c:\users\user\appdata\local\programs\python\python39\lib\site-packages (from torch==2.7.0.dev20250311) (1.13.3) Requirement already satisfied: networkx in c:\users\user\appdata\local\programs\python\python39\lib\site-packages (from torch==2.7.0.dev20250311) (3.2.1) Requirement already satisfied: jinja2 in c:\users\user\appdata\local\programs\python\python39\lib\site-packages (from torch==2.7.0.dev20250311) (3.1.4) Requirement already satisfied: fsspec in c:\users\user\appdata\local\programs\python\python39\lib\site-packages (from torch==2.7.0.dev20250311) (2024.10.0) Requirement already satisfied: numpy in c:\users\user\appdata\local\programs\python\python39\lib\site-packages (from torchvision==0.22.0.dev20250312+cu128) (1.26.4) Requirement already satisfied: pillow!=8.3.*,>=5.3.0 in c:\users\user\appdata\local\programs\python\python39\lib\site-packages (from torchvision==0.22.0.dev20250312+cu128) (11.0.0) Requirement already satisfied: mpmath<1.4,>=1.1.0 in c:\users\user\appdata\local\programs\python\python39\lib\site-packages (from sympy>=1.13.3->torch==2.7.0.dev20250311) (1.3.0) Requirement already satisfied: MarkupSafe>=2.0 in c:\users\user\appdata\local\programs\python\python39\lib\site-packages (from jinja2->torch==2.7.0.dev20250311) (2.1.5) ERROR: flash_attn-2.7.4.post1-cp310-cp310-win_amd64.whl is not a supported wheel on this platform. ERROR: deepspeed-0.16.4-cp310-cp310-win_amd64.whl is not a supported wheel on this platform. ERROR: sageattention-2.1.1+cu128torch2.7.0-cp310-cp310-win_amd64.whl is not a supported wheel on this platform. Requirement already satisfied: triton-windows in c:\users\user\appdata\local\programs\python\python39\lib\site-packages (3.3.0a0.post18) Requirement already satisfied: setuptools>=40.8.0 in c:\users\user\appdata\local\programs\python\python39\lib\site-packages (from triton-windows) (58.1.0) Requirement already satisfied: torchao in c:\users\user\appdata\local\programs\python\python39\lib\site-packages (0.10.0) ERROR: Could not open requirements file: [Errno 2] No such file or directory: 'requirements.txt' Requirement already satisfied: huggingface_hub in c:\users\user\appdata\local\programs\python\python39\lib\site-packages (0.30.2) Requirement already satisfied: ipywidgets in c:\users\user\appdata\local\programs\python\python39\lib\site-packages (8.1.6) Requirement already satisfied: hf_transfer in c:\users\user\appdata\local\programs\python\python39\lib\site-packages (0.1.9) Requirement already satisfied: filelock in c:\users\user\appdata\local\programs\python\python39\lib\site-packages (from huggingface_hub) (3.16.1) Requirement already satisfied: fsspec>=2023.5.0 in c:\users\user\appdata\local\programs\python\python39\lib\site-packages (from huggingface_hub) (2024.10.0) Requirement already satisfied: packaging>=20.9 in c:\users\user\appdata\local\programs\python\python39\lib\site-packages (from huggingface_hub) (24.2) Requirement already satisfied: pyyaml>=5.1 in c:\users\user\appdata\local\programs\python\python39\lib\site-packages (from huggingface_hub) (6.0.2) Requirement already satisfied: requests in c:\users\user\appdata\local\programs\python\python39\lib\site-packages (from huggingface_hub) (2.32.3) Requirement already satisfied: tqdm>=4.42.1 in c:\users\user\appdata\local\programs\python\python39\lib\site-packages (from huggingface_hub) (4.67.1) Requirement already satisfied: typing-extensions>=3.7.4.3 in c:\users\user\appdata\local\programs\python\python39\lib\site-packages (from huggingface_hub) (4.13.0) Requirement already satisfied: comm>=0.1.3 in c:\users\user\appdata\local\programs\python\python39\lib\site-packages (from ipywidgets) (0.2.2) Requirement already satisfied: ipython>=6.1.0 in c:\users\user\appdata\local\programs\python\python39\lib\site-packages (from ipywidgets) (8.18.1) Requirement already satisfied: traitlets>=4.3.1 in c:\users\user\appdata\local\programs\python\python39\lib\site-packages (from ipywidgets) (5.14.3) Requirement already satisfied: widgetsnbextension~=4.0.14 in c:\users\user\appdata\local\programs\python\python39\lib\site-packages (from ipywidgets) (4.0.14) Requirement already satisfied: jupyterlab_widgets~=3.0.14 in c:\users\user\appdata\local\programs\python\python39\lib\site-packages (from ipywidgets) (3.0.14) Requirement already satisfied: decorator in c:\users\user\appdata\local\programs\python\python39\lib\site-packages (from ipython>=6.1.0->ipywidgets) (5.2.1) Requirement already satisfied: jedi>=0.16 in c:\users\user\appdata\local\programs\python\python39\lib\site-packages (from ipython>=6.1.0->ipywidgets) (0.19.2) Requirement already satisfied: matplotlib-inline in c:\users\user\appdata\local\programs\python\python39\lib\site-packages (from ipython>=6.1.0->ipywidgets) (0.1.7) Requirement already satisfied: prompt-toolkit<3.1.0,>=3.0.41 in c:\users\user\appdata\local\programs\python\python39\lib\site-packages (from ipython>=6.1.0->ipywidgets) (3.0.51) Requirement already satisfied: pygments>=2.4.0 in c:\users\user\appdata\local\programs\python\python39\lib\site-packages (from ipython>=6.1.0->ipywidgets) (2.19.1) Requirement already satisfied: stack-data in c:\users\user\appdata\local\programs\python\python39\lib\site-packages (from ipython>=6.1.0->ipywidgets) (0.6.3) Requirement already satisfied: exceptiongroup in c:\users\user\appdata\local\programs\python\python39\lib\site-packages (from ipython>=6.1.0->ipywidgets) (1.2.2) Requirement already satisfied: colorama in c:\users\user\appdata\local\programs\python\python39\lib\site-packages (from ipython>=6.1.0->ipywidgets) (0.4.6) Requirement already satisfied: charset-normalizer<4,>=2 in c:\users\user\appdata\local\programs\python\python39\lib\site-packages (from requests->huggingface_hub) (3.4.1) Requirement already satisfied: idna<4,>=2.5 in c:\users\user\appdata\local\programs\python\python39\lib\site-packages (from requests->huggingface_hub) (3.10) Requirement already satisfied: urllib3<3,>=1.21.1 in c:\users\user\appdata\local\programs\python\python39\lib\site-packages (from requests->huggingface_hub) (2.3.0) Requirement already satisfied: certifi>=2017.4.17 in c:\users\user\appdata\local\programs\python\python39\lib\site-packages (from requests->huggingface_hub) (2025.1.31) Requirement already satisfied: parso<0.9.0,>=0.8.4 in c:\users\user\appdata\local\programs\python\python39\lib\site-packages (from jedi>=0.16->ipython>=6.1.0->ipywidgets) (0.8.4) Requirement already satisfied: wcwidth in c:\users\user\appdata\local\programs\python\python39\lib\site-packages (from prompt-toolkit<3.1.0,>=3.0.41->ipython>=6.1.0->ipywidgets) (0.2.13) Requirement already satisfied: executing>=1.2.0 in c:\users\user\appdata\local\programs\python\python39\lib\site-packages (from stack-data->ipython>=6.1.0->ipywidgets) (2.2.0) Requirement already satisfied: asttokens>=2.1.0 in c:\users\user\appdata\local\programs\python\python39\lib\site-packages (from stack-data->ipython>=6.1.0->ipywidgets) (3.0.0) Requirement already satisfied: pure-eval in c:\users\user\appdata\local\programs\python\python39\lib\site-packages (from stack-data->ipython>=6.1.0->ipywidgets) (0.2.3) 'git' is not recognized as an internal or external command, operable program or batch file. 'git' is not recognized as an internal or external command, operable program or batch file. ERROR: Could not open requirements file: [Errno 2] No such file or directory: 'requirements.txt' Fetching 10 files: 100%|█████████████████████████████████████████████████████████████| 10/10 [00:00<00:00, 1295.78it/s] FramePack installed check out all messages and save them before close to verify any errors or not later Press any key to continue . . . and after open windows app start nothing happens

Damien Jourdan

my suggestion is to get a video from A to B, then make another video from B to A without reloading the pictures. I use to paste them to do a loop but with having the control of the middle frame and eventually double the duration. If it can be done directly from your (amazing) tool, it would be perfect.

SECourses

so you want a button to swap uploaded pictures right?

SECourses

thanks i found the error working on fix

TaTaKae

It’s the same with all LoRAs, especially since you mentioned they use an extra 20GB of VRAM. I need at least an RTX 5090 with 32GB of VRAM to get it working on RunPod, and even then, I sometimes run into out-of-memory errors. Also, it seems like the LoRAs only work at the first generation. After that, they stop working completely unless I restart Framepack. I've tested this with over 10 different LoRAs, and it’s always the same issue — they only work for the first generation, and then nothing. I have no idea why this is happening.

W4v3r1d3r

Hi, I'm getting an error that the app cannot save a file due to an error. It happened in a masscompute session and also on my local PC (RTX4080). The final mp4 is created, but not the processing metadata and the last frame of generations (both checkbox are selected) this is the last output Offloading DynamicSwap_HunyuanVideoTransformer3DModelPacked from cuda:0 to preserve memory: 8 GB Loaded AutoencoderKLHunyuanVideo to cuda:0 as complete. VAE decoding completed in 23.21 seconds (0.029 sec/frame) Unloaded AutoencoderKLHunyuanVideo as complete. Saved MP4 video to /home/Ubuntu/Downloads/FramePack_v30/FramePack/outputs/2025_04_23_11_12_28_097.mp4 Error saving MP4/WebM video: 'Checkbox' object is not callable Saving operations completed in 71.59 seconds Decoded. Current latent shape torch.Size([1, 16, 793, 126, 84]); pixel shape torch.Size([1, 3, 3169, 1008, 672]) Generation 1/1 completed in 13634.74 seconds Total generation time: 13634.74 seconds (227.25 minutes) Warning: No output file was generated due to an error

SECourses

hello i am about to fix it sorry for delay. fixing the last errors. sorry for this. wait me to update next

SECourses

please download v31 and use it fixed. ty

SECourses

please download v31 and use it fixed. ty

SECourses

please download v31 and use it fixed. ty

SECourses

please download v31 and use it fixed. ty

W4v3r1d3r

No problem, just reporting. You are doing an amazing work!

koon grizzly

i was testing v31 with 15fps and 2x rife and 10 fps and 4x rife but i don't notice much difference between original and extra fps version.? I can add that the prompt with seconds works now

SECourses

awesome. i am improving second thing. with v33 you will see Latent Window Size . it is not perfected yet but i will add a button to distribute app to generate every 1 second splits. Latent Window Size impacts how many processing it takes by the way v33 published

Mark Sutherland

Fantastic work as always, glad you got your github sorted. Will you be releasing a one click installer for instantcharacter, looks like it has alot of potential especially if you can add more loras, cheers

koon grizzly

nice, you release updates faster then i am able to test haha, meanwhile my next attempt with the second prompts was less succesfull, i noticed in the info while processing it kept showing same 'current prompt' so it did only one thing in 10 seconds (v31, i will now install new version)

stphnvdb321

I also tested 10 fps and Rife 4x. It gives me a slow motion, even if it is a little faster.

SECourses

Great let me know how it goes last version. Still some minor problems

SECourses

Yes 10 fps literally being slow motion since original model is 30 fps :)

Ec Jep

versions 25 and prior worked but I get this after a fresh install of v31: Edit this file and set set HF_HUB_ENABLE_HF_TRANSFER=0 if you get model download errors of HF_TRANSFER Hugging Face cache folders set to: D:\SD\framepack_v31\FramePack\models W0423 13:34:47.665000 19836 venv\Lib\site-packages\torch\distributed\elastic\multiprocessing\redirects.py:29] NOTE: Redirects are currently not supported in Windows or MacOs. Currently enabled native sdp backends: ['flash', 'math', 'mem_efficient', 'cudnn'] Xformers is not installed! Flash Attn is installed! Sage Attn is installed! Namespace(share=False) Free VRAM 22.455078125 GB High-VRAM Mode: False Downloading shards: 100%|██████████████████████████████████████████████████████████████| 4/4 [00:00<00:00, 1275.64it/s] Loading checkpoint shards: 100%|█████████████████████████████████████████████████████████| 4/4 [00:00<00:00, 8.37it/s] Traceback (most recent call last): File "D:\SD\framepack_v31\FramePack\venv\lib\site-packages\diffusers\models\model_loading_utils.py", line 176, in load_state_dict return safetensors.torch.load_file(checkpoint_file, device=map_location) File "D:\SD\framepack_v31\FramePack\venv\lib\site-packages\safetensors\torch.py", line 313, in load_file with safe_open(filename, framework="pt", device=device) as f: safetensors_rust.SafetensorError: Error while deserializing header: InvalidHeaderDeserialization During handling of the above exception, another exception occurred: Traceback (most recent call last): File "D:\SD\framepack_v31\FramePack\venv\lib\site-packages\diffusers\models\model_loading_utils.py", line 195, in load_state_dict if f.read().startswith("version"): File "C:\Users\mtnme\AppData\Local\Programs\Python\Python310\lib\encodings\cp1252.py", line 23, in decode return codecs.charmap_decode(input,self.errors,decoding_table)[0] UnicodeDecodeError: 'charmap' codec can't decode byte 0x8f in position 985661474: character maps to During handling of the above exception, another exception occurred: Traceback (most recent call last): File "D:\SD\framepack_v31\FramePack\app.py", line 59, in vae = AutoencoderKLHunyuanVideo.from_pretrained("hunyuanvideo-community/HunyuanVideo", subfolder='vae', torch_dtype=torch.float16).cpu() File "D:\SD\framepack_v31\FramePack\venv\lib\site-packages\huggingface_hub\utils\_validators.py", line 114, in _inner_fn return fn(*args, **kwargs) File "D:\SD\framepack_v31\FramePack\venv\lib\site-packages\diffusers\models\modeling_utils.py", line 1190, in from_pretrained state_dict = load_state_dict(resolved_model_file[0], disable_mmap=disable_mmap, dduf_entries=dduf_entries) File "D:\SD\framepack_v31\FramePack\venv\lib\site-packages\diffusers\models\model_loading_utils.py", line 207, in load_state_dict raise OSError( OSError: Unable to load weights from checkpoint file for 'D:\SD\framepack_v31\FramePack\hf_download\hub\models--hunyuanvideo-community--HunyuanVideo\snapshots\e8c2aaa66fe3742a32c11a6766aecbf07c56e773\vae\diffusion_pytorch_model.safetensors' at 'D:\SD\framepack_v31\FramePack\hf_download\hub\models--hunyuanvideo-community--HunyuanVideo\snapshots\e8c2aaa66fe3742a32c11a6766aecbf07c56e773\vae\diffusion_pytorch_model.safetensors'. Press any key to continue . . .

SECourses

Looks like models corrupted somehow. Can you install zip file into a new folder and let me know? Use latest version

koon grizzly

For me v35 still gives same video on 2x and 4x rife compared to original video. (only change is file size is smaller) . I gave up at the moment with the seconds prompting, i have more success not writing any prompt at all, create 5 second videos and use last frame to create the next video to be able to get some movement at least once every 5 seconds and if it doesn't it's only 5 seconds that need to be re created :-)

Dan

Hi, thank you for this great app. I am wondering about the LoRA option. What kind of LoRA can one use? Is there a collection of LoRA available somewhere? Thank you.

SECourses

please try v37 and set Set Window for ~1s Sections and let me know. now way better

macmotu

I love it. It get´s better every day! Thank you very much.

SECourses

you are welcome and thanks for support

Thomas McGrath

Can this workflow be modified to support WAN 2.1 as well?

Dan

ok, cool. Thank you for the honest answer. I understand this is all so experimental, and you are relentlessly updating everything! This is so valuable! This app is great, keep up the great work, so much fun!

SECourses

not yet sadly. i am following developments. but for wan 2.1 hopefully i will bring begin and ending images soon

SECourses

thanks i am still in research will update the app when i got better solution

RenderDrgn

Hello just a quick question. Is there a way to run this without an internet connection?

Jesen

Is there a way to downgrade to an earlier version without deleting and reinstalling the whole thing? The v40 version does not seem to utilize system virtual ram like v21 did

Male AI Art

The video seems to be generated without problems, but I get this message. Is there a problem? Total worker time: 669.96 seconds (11.17 minutes) ERROR:asyncio:Exception in callback _ProactorBasePipeTransport._call_connection_lost(None) handle: Traceback (most recent call last): File "C:\Users\****\AppData\Local\Programs\Python\Python310\lib\asyncio\events.py", line 80, in _run self._context.run(self._callback, *self._args) File "C:\Users\****\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 ERROR:asyncio:Exception in callback _ProactorBasePipeTransport._call_connection_lost(None) handle: Traceback (most recent call last): File "C:\Users\****\AppData\Local\Programs\Python\Python310\lib\asyncio\events.py", line 80, in _run self._context.run(self._callback, *self._args) File "C:\Users\****\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.

Weronika

I have a problem: after the first launch and successful generation, the app won't start again. What am I doing wrong? Edit this file and set set HF_HUB_ENABLE_HF_TRANSFER=0 if you get model download errors of HF_TRANSFER Hugging Face cache folders set to: C:\FramePack\models W0424 07:03:32.385865 12972 venv\Lib\site-packages\torch\distributed\elastic\multiprocessing\redirects.py:29] NOTE: Redirects are currently not supported in Windows or MacOs. Currently enabled native sdp backends: ['flash', 'math', 'mem_efficient', 'cudnn'] Xformers is not installed! Flash Attn is installed! Sage Attn is installed! Namespace(share=False) Free VRAM 4.9453544318675995 GB High-VRAM Mode: False Downloading shards: 100%|██████████████████████████████████████████████████████████████| 4/4 [00:00<00:00, 2428.66it/s] Press any key to continue . . .

SECourses

hi it is out of RAM. please set your virtual RAM to 50 GB or above

SECourses

No this is Gradio related error. you can ignore. also i would install my python into c drive as in tutorial video

SECourses

can you elaborate more? the only change regarding virtual ram is that vae decoding was hard coded to 8 gb. now it will use your Preserved Memory parameter. how much did you set?

Rudbby

Cool thank you, switched over to massed compute and it worked like a charm.

Taiga

when I try to use a lora with it, I get errors and it quits before getting started. I have a 24BG 4090: File "D:\FramePack\FramePack\app.py", line 965, in worker transformer.to(gpu) File "D:\FramePack\FramePack\venv\lib\site-packages\diffusers\models\modeling_utils.py", line 1353, in to return super().to(*args, **kwargs) File "D:\FramePack\FramePack\venv\lib\site-packages\torch\nn\modules\module.py", line 1355, in to return self._apply(convert) File "D:\FramePack\FramePack\venv\lib\site-packages\torch\nn\modules\module.py", line 915, in _apply module._apply(fn) File "D:\FramePack\FramePack\venv\lib\site-packages\torch\nn\modules\module.py", line 915, in _apply module._apply(fn) File "D:\FramePack\FramePack\venv\lib\site-packages\torch\nn\modules\module.py", line 915, in _apply module._apply(fn) File "D:\FramePack\FramePack\venv\lib\site-packages\torch\nn\modules\module.py", line 942, in _apply param_applied = fn(param) File "D:\FramePack\FramePack\venv\lib\site-packages\torch\nn\modules\module.py", line 1341, in convert return t.to( torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 72.00 MiB. GPU 0 has a total capacity of 22.49 GiB of which 0 bytes is free. Of the allocated memory 21.48 GiB is allocated by PyTorch, and 24.87 MiB 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) latent_padding_size = 9, is_last_section = False, is_first_section = True Unloaded SiglipVisionModel as complete. Moving DynamicSwap_HunyuanVideoTransformer3DModelPacked to cuda:0 with preserved memory: 8 GB Teacache: 0.15

KonekoFuta

Hey, what I've seen so far seems great. It doesn't seem like batch generation is working. It stops after the first set of generations that used the first image in the batch input folder. Is that expected?

SECourses

can you email me your exact configuration? batch should work perfectly fine. monstermmorpg@gmail.com - if broken will fix asap hopefully

stphnvdb321

I just ran the installer again to upgrade to V40 and got the following error message: "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. moviepy 2.1.2 requires pillow<11.0,>=9.2.0, but you have pillow 11.1.0 which is incompatible." Do I have to worry?

stphnvdb321

Yes, FramePack with Wan2.1 would be awesome. From my eyperience, Hunyuan is much worse to follow a promt. That't the why I uninstalled Hunyuan again.

koon grizzly

Thanks, just got back from work, meanwhile it's v40 already lol

Thomas McGrath

Check out the Pinokio WAN 2.1 app. They have start/end images using various WAN models. But they do not have the wonderful image saving, preview viewing that you app has. Those features save vast amounts of time as you can evaluate/terminate the process without having to wait for the final video.

SECourses

yes i will add begin end frame very soon hopefully. currently finishing the framepack app

koon grizzly

Just created 2x a 10 second video where it mostly follows the timestamp prompts, woohoo, let's hope the videos where nothing happens are a thing of the past now

Hockey

Would it be possible to add llava for describing the input image. Then we can use the llava prompt to change or add to it.

TAGopolis

I can run this once, then before running it again I have to reboot my machine or it errors out with this error: Exception Code: 0xC0000005 0x00007FFAE06E60B4, D:\AIServer\FramePack\FramePack\venv\lib\site-packages\torch\lib\c10.dll(0x00007FFAE0660000) + 0x860B4 byte(s), ?typeMetaDatas@TypeMeta@caffe2@@CAPEAUTypeMetaData@detail@2@XZ() + 0x6E14 byte(s) 0x00007FFAE06E54EC, D:\AIServer\FramePack\FramePack\venv\lib\site-packages\torch\lib\c10.dll(0x00007FFAE0660000) + 0x854EC byte(s), ?typeMetaDatas@TypeMeta@caffe2@@CAPEAUTypeMetaData@detail@2@XZ() + 0x624C byte(s) 0x00007FFAE06E7875, D:\AIServer\FramePack\FramePack\venv\lib\site-packages\torch\lib\c10.dll(0x00007FFAE0660000) + 0x87875 byte(s), ?typeMetaDatas@TypeMeta@caffe2@@CAPEAUTypeMetaData@detail@2@XZ() + 0x85D5 byte(s) 0x00007FFAE06E7513, D:\AIServer\FramePack\FramePack\venv\lib\site-packages\torch\lib\c10.dll(0x00007FFAE0660000) + 0x87513 byte(s), ?typeMetaDatas@TypeMeta@caffe2@@CAPEAUTypeMetaData@detail@2@XZ() + 0x8273 byte(s) 0x00007FFAE06E6CB0, D:\AIServer\FramePack\FramePack\venv\lib\site-packages\torch\lib\c10.dll(0x00007FFAE0660000) + 0x86CB0 byte(s), ?typeMetaDatas@TypeMeta@caffe2@@CAPEAUTypeMetaData@detail@2@XZ() + 0x7A10 byte(s) 0x00007FFAE06E437E, D:\AIServer\FramePack\FramePack\venv\lib\site-packages\torch\lib\c10.dll(0x00007FFAE0660000) + 0x8437E byte(s), ?typeMetaDatas@TypeMeta@caffe2@@CAPEAUTypeMetaData@detail@2@XZ() + 0x50DE byte(s) 0x00007FFAE06E4429, D:\AIServer\FramePack\FramePack\venv\lib\site-packages\torch\lib\c10.dll(0x00007FFAE0660000) + 0x84429 byte(s), ?typeMetaDatas@TypeMeta@caffe2@@CAPEAUTypeMetaData@detail@2@XZ() + 0x5189 byte(s) 0x00007FFAE06B5138, D:\AIServer\FramePack\FramePack\venv\lib\site-packages\torch\lib\c10.dll(0x00007FFAE0660000) + 0x55138 byte(s), ?alloc_cpu@c10@@YAPEAX_K@Z() + 0x58 byte(s) 0x00007FFAE066963C, D:\AIServer\FramePack\FramePack\venv\lib\site-packages\torch\lib\c10.dll(0x00007FFAE0660000) + 0x963C byte(s), ?allocate@DefaultCPUAllocator@c10@@UEAA?AVDataPtr@2@_K@Z() + 0x3C byte(s) 0x00007FFAE06838D3, D:\AIServer\FramePack\FramePack\venv\lib\site-packages\torch\lib\c10.dll(0x00007FFAE0660000) + 0x238D3 byte(s), ??0StorageImpl@c10@@QEAA@Uuse_byte_size_t@01@AEBVSymInt@1@PEAUAllocator@1@_N@Z() + 0x63 byte(s) 0x00007FF9B9E47349, D:\AIServer\FramePack\FramePack\venv\lib\site-packages\torch\lib\torch_cpu.dll(0x00007FF9B4370000) + 0x5AD7349 byte(s), ?sym@DynamicLibrary@at@@QEAAPEAXPEBD@Z() + 0xD39 byte(s) 0x00007FF9B9E49577, D:\AIServer\FramePack\FramePack\venv\lib\site-packages\torch\lib\torch_cpu.dll(0x00007FF9B4370000) + 0x5AD9577 byte(s), ?empty_strided_cpu@detail@at@@YA?AVTensorBase@2@V?$ArrayRef@_J@c10@@0V?$optional@W4ScalarType@c10@@@std@@V?$optional@W4Layout@c10@@@7@V?$optional@UDevice@c10@@@7@V?$optional@_N@7@@Z() + 0x97 byte(s) 0x00007FF9BA3B8C09, D:\AIServer\FramePack\FramePack\venv\lib\site-packages\torch\lib\torch_cpu.dll(0x00007FF9B4370000) + 0x6048C09 byte(s), ?empty_strided_cpu@native@at@@YA?AVTensor@2@V?$ArrayRef@_J@c10@@0V?$optional@W4ScalarType@c10@@@std@@V?$optional@W4Layout@c10@@@7@V?$optional@UDevice@c10@@@7@V?$optional@_N@7@@Z() + 0x79 byte(s) 0x00007FF9BAEE1190, D:\AIServer\FramePack\FramePack\venv\lib\site-packages\torch\lib\torch_cpu.dll(0x00007FF9B4370000) + 0x6B71190 byte(s), ?where_outf@cpu@at@@YAAEAVTensor@2@AEBV32@00AEAV32@@Z() + 0x960 byte(s) 0x00007FF9BAED8AF8, D:\AIServer\FramePack\FramePack\venv\lib\site-packages\torch\lib\torch_cpu.dll(0x00007FF9B4370000) + 0x6B68AF8 byte(s), ?bucketize_outf@cpu@at@@YAAEAVTensor@2@AEBV32@0_N1AEAV32@@Z() + 0xFDE8 byte(s) 0x00007FF9BAA2A905, D:\AIServer\FramePack\FramePack\venv\lib\site-packages\torch\lib\torch_cpu.dll(0x00007FF9B4370000) + 0x66BA905 byte(s), ?redispatch@xlogy__Tensor@_ops@at@@SAAEAVTensor@3@VDispatchKeySet@c10@@AEAV43@AEBV43@@Z() + 0x1A665 byte(s) 0x00007FF9BAA9B51B, D:\AIServer\FramePack\FramePack\venv\lib\site-packages\torch\lib\torch_cpu.dll(0x00007FF9B4370000) + 0x672B51B byte(s), ?redispatch@xlogy__Tensor@_ops@at@@SAAEAVTensor@3@VDispatchKeySet@c10@@AEAV43@AEBV43@@Z() + 0x8B27B byte(s) 0x00007FF9BAB79A3E, D:\AIServer\FramePack\FramePack\venv\lib\site-packages\torch\lib\torch_cpu.dll(0x00007FF9B4370000) + 0x6809A3E byte(s), ?redispatch@empty_strided@_ops@at@@SA?AVTensor@3@VDispatchKeySet@c10@@V?$ArrayRef@VSymInt@c10@@@6@1V?$optional@W4ScalarType@c10@@@std@@V?$optional@W4Layout@c10@@@9@V?$optional@UDevice@c10@@@9@V?$optional@_N@9@@Z() + 0xEE byte(s) 0x00007FF9BAE071FF, D:\AIServer\FramePack\FramePack\venv\lib\site-packages\torch\lib\torch_cpu.dll(0x00007FF9B4370000) + 0x6A971FF byte(s), ?redispatch@view_as_real@_ops@at@@SA?AVTensor@3@VDispatchKeySet@c10@@AEBV43@@Z() + 0x2407F byte(s) 0x00007FF9BAE04008, D:\AIServer\FramePack\FramePack\venv\lib\site-packages\torch\lib\torch_cpu.dll(0x00007FF9B4370000) + 0x6A94008 byte(s), ?redispatch@view_as_real@_ops@at@@SA?AVTensor@3@VDispatchKeySet@c10@@AEBV43@@Z() + 0x20E88 byte(s) 0x00007FF9BAA2A905, D:\AIServer\FramePack\FramePack\venv\lib\site-packages\torch\lib\torch_cpu.dll(0x00007FF9B4370000) + 0x66BA905 byte(s), ?redispatch@xlogy__Tensor@_ops@at@@SAAEAVTensor@3@VDispatchKeySet@c10@@AEAV43@AEBV43@@Z() + 0x1A665 byte(s) 0x00007FF9BAAE57E0, D:\AIServer\FramePack\FramePack\venv\lib\site-packages\torch\lib\torch_cpu.dll(0x00007FF9B4370000) + 0x67757E0 byte(s), ?call@empty_strided@_ops@at@@SA?AVTensor@3@V?$ArrayRef@VSymInt@c10@@@c10@@0V?$optional@W4ScalarType@c10@@@std@@V?$optional@W4Layout@c10@@@8@V?$optional@UDevice@c10@@@8@V?$optional@_N@8@@Z() + 0x260 byte(s) 0x00007FF9B9EAFF28, D:\AIServer\FramePack\FramePack\venv\lib\site-packages\torch\lib\torch_cpu.dll(0x00007FF9B4370000) + 0x5B3FF28 byte(s), ?declare_static_shape@TensorIteratorConfig@at@@QEAAAEAV12@V?$ArrayRef@_J@c10@@@Z() + 0x3A8 byte(s) 0x00007FF9BA3A4892, D:\AIServer\FramePack\FramePack\venv\lib\site-packages\torch\lib\torch_cpu.dll(0x00007FF9B4370000) + 0x6034892 byte(s), ?_to_copy@native@at@@YA?AVTensor@2@AEBV32@V?$optional@W4ScalarType@c10@@@std@@V?$optional@W4Layout@c10@@@5@V?$optional@UDevice@c10@@@5@V?$optional@_N@5@_NV?$optional@W4MemoryFormat@c10@@@5@@Z() + 0x822 byte(s) 0x00007FF9BB11B89C, D:\AIServer\FramePack\FramePack\venv\lib\site-packages\torch\lib\torch_cpu.dll(0x00007FF9B4370000) + 0x6DAB89C byte(s), ?view_copy_symint_outf@compositeexplicitautograd@at@@YAAEAVTensor@2@AEBV32@V?$ArrayRef@VSymInt@c10@@@c10@@AEAV32@@Z() + 0x2FDC byte(s) 0x00007FF9BB0EBB82, D:\AIServer\FramePack\FramePack\venv\lib\site-packages\torch\lib\torch_cpu.dll(0x00007FF9B4370000) + 0x6D7BB82 byte(s), ?bucketize_outf@compositeexplicitautograd@at@@YAAEAVTensor@2@AEBVScalar@c10@@AEBV32@_N2AEAV32@@Z() + 0x478F2 byte(s) 0x00007FF9BA86DDC8, D:\AIServer\FramePack\FramePack\venv\lib\site-packages\torch\lib\torch_cpu.dll(0x00007FF9B4370000) + 0x64FDDC8 byte(s), ?redispatch@_to_copy@_ops@at@@SA?AVTensor@3@VDispatchKeySet@c10@@AEBV43@V?$optional@W4ScalarType@c10@@@std@@V?$optional@W4Layout@c10@@@8@V?$optional@UDevice@c10@@@8@V?$optional@_N@8@_NV?$optional@W4MemoryFormat@c10@@@8@@Z() + 0xE8 byte(s) 0x00007FF9BADFE802, D:\AIServer\FramePack\FramePack\venv\lib\site-packages\torch\lib\torch_cpu.dll(0x00007FF9B4370000) + 0x6A8E802 byte(s), ?redispatch@view_as_real@_ops@at@@SA?AVTensor@3@VDispatchKeySet@c10@@AEBV43@@Z() + 0x1B682 byte(s) 0x00007FF9BAE038A2, D:\AIServer\FramePack\FramePack\venv\lib\site-packages\torch\lib\torch_cpu.dll(0x00007FF9B4370000) + 0x6A938A2 byte(s), ?redispatch@view_as_real@_ops@at@@SA?AVTensor@3@VDispatchKeySet@c10@@AEBV43@@Z() + 0x20722 byte(s) 0x00007FF9BA86DDC8, D:\AIServer\FramePack\FramePack\venv\lib\site-packages\torch\lib\torch_cpu.dll(0x00007FF9B4370000) + 0x64FDDC8 byte(s), ?redispatch@_to_copy@_ops@at@@SA?AVTensor@3@VDispatchKeySet@c10@@AEBV43@V?$optional@W4ScalarType@c10@@@std@@V?$optional@W4Layout@c10@@@8@V?$optional@UDevice@c10@@@8@V?$optional@_N@8@_NV?$optional@W4MemoryFormat@c10@@@8@@Z() + 0xE8 byte(s) 0x00007FF9BC3A2C6A, D:\AIServer\FramePack\FramePack\venv\lib\site-packages\torch\lib\torch_cpu.dll(0x00007FF9B4370000) + 0x8032C6A byte(s), ??1NotImplemented@autograd@torch@@UEAA@XZ() + 0x4CF1A byte(s) 0x00007FF9BC3C8001, D:\AIServer\FramePack\FramePack\venv\lib\site-packages\torch\lib\torch_cpu.dll(0x00007FF9B4370000) + 0x8058001 byte(s), ?apply_with_saved@UndefinedGradBackward@autograd@torch@@UEAA?AV?$vector@VTensor@at@@V?$allocator@VTensor@at@@@std@@@std@@AEBV45@AEAVSwapSavedVariables@2dynamo@3@@Z() + 0x1B011 byte(s) 0x00007FF9BA7DB513, D:\AIServer\FramePack\FramePack\venv\lib\site-packages\torch\lib\torch_cpu.dll(0x00007FF9B4370000) + 0x646B513 byte(s), ?call@_to_copy@_ops@at@@SA?AVTensor@3@AEBV43@V?$optional@W4ScalarType@c10@@@std@@V?$optional@W4Layout@c10@@@6@V?$optional@UDevice@c10@@@6@V?$optional@_N@6@_NV?$optional@W4MemoryFormat@c10@@@6@@Z() + 0x203 byte(s) 0x00007FF9BA3AD3CE, D:\AIServer\FramePack\FramePack\venv\lib\site-packages\torch\lib\torch_cpu.dll(0x00007FF9B4370000) + 0x603D3CE byte(s), ?to_dense_backward@native@at@@YA?AVTensor@2@AEBV32@0V?$optional@_N@std@@@Z() + 0x54E byte(s) 0x00007FF9BA3ACC57, D:\AIServer\FramePack\FramePack\venv\lib\site-packages\torch\lib\torch_cpu.dll(0x00007FF9B4370000) + 0x603CC57 byte(s), ?to@native@at@@YA?AVTensor@2@AEBV32@V?$optional@W4ScalarType@c10@@@std@@V?$optional@W4Layout@c10@@@5@V?$optional@UDevice@c10@@@5@V?$optional@_N@5@_N5V?$optional@W4MemoryFormat@c10@@@5@@Z() + 0x107 byte(s) 0x00007FF9BB1CF658, D:\AIServer\FramePack\FramePack\venv\lib\site-packages\torch\lib\torch_cpu.dll(0x00007FF9B4370000) + 0x6E5F658 byte(s), ?where@compositeimplicitautograd@at@@YA?AVTensor@2@AEBV32@AEBVScalar@c10@@1@Z() + 0x7148 byte(s) 0x00007FF9BB1B26EE, D:\AIServer\FramePack\FramePack\venv\lib\site-packages\torch\lib\torch_cpu.dll(0x00007FF9B4370000) + 0x6E426EE byte(s), ?broadcast_to_symint@compositeimplicitautograd@at@@YA?AVTensor@2@AEBV32@V?$ArrayRef@VSymInt@c10@@@c10@@@Z() + 0x3AA6E byte(s) 0x00007FF9BA8B9675, D:\AIServer\FramePack\FramePack\venv\lib\site-packages\torch\lib\torch_cpu.dll(0x00007FF9B4370000) + 0x6549675 byte(s), ?redispatch@zeros_out@_ops@at@@SAAEAVTensor@3@VDispatchKeySet@c10@@V?$ArrayRef@VSymInt@c10@@@6@AEAV43@@Z() + 0x17EF5 byte(s) 0x00007FF9BA9AB6FE, D:\AIServer\FramePack\FramePack\venv\lib\site-packages\torch\lib\torch_cpu.dll(0x00007FF9B4370000) + 0x663B6FE byte(s), ?call@to_dtype_layout@_ops@at@@SA?AVTensor@3@AEBV43@V?$optional@W4ScalarType@c10@@@std@@V?$optional@W4Layout@c10@@@6@V?$optional@UDevice@c10@@@6@V?$optional@_N@6@_N5V?$optional@W4MemoryFormat@c10@@@6@@Z() + 0x26E byte(s) 0x00007FF9B9DF0732, D:\AIServer\FramePack\FramePack\venv\lib\site-packages\torch\lib\torch_cpu.dll(0x00007FF9B4370000) + 0x5A80732 byte(s), ?to@Tensor@at@@QEBA?AV12@UTensorOptions@c10@@_N1V?$optional@W4MemoryFormat@c10@@@std@@@Z() + 0x1D2 byte(s) 0x00007FF9B30F7387, D:\AIServer\FramePack\FramePack\venv\lib\site-packages\torch\lib\torch_python.dll(0x00007FF9B2FD0000) + 0x127387 byte(s), ?dimnamelist@PythonArgs@torch@@QEAA?AV?$vector@UDimname@at@@V?$allocator@UDimname@at@@@std@@@std@@H@Z() + 0x387 byte(s) 0x00007FF9B30E4598, D:\AIServer\FramePack\FramePack\venv\lib\site-packages\torch\lib\torch_python.dll(0x00007FF9B2FD0000) + 0x114598 byte(s), ?get@?$THPPointer@UPyCodeObject@@@@QEAAPEAUPyCodeObject@@XZ() + 0xCE3E8 byte(s) 0x00007FFAE9DFEE6F, C:\Python310\python310.dll(0x00007FFAE9DA0000) + 0x5EE6F byte(s), PyObject_SetAttr() + 0xC57 byte(s) 0x00007FFAE9E16385, C:\Python310\python310.dll(0x00007FFAE9DA0000) + 0x76385 byte(s), _PyEval_EvalFrameDefault() + 0xAE5 byte(s) 0x00007FFAE9E149D7, C:\Python310\python310.dll(0x00007FFAE9DA0000) + 0x749D7 byte(s), _PyFunction_Vectorcall() + 0x87 byte(s) 0x00007FFAE9E16033, C:\Python310\python310.dll(0x00007FFAE9DA0000) + 0x76033 byte(s), _PyEval_EvalFrameDefault() + 0x793 byte(s) 0x00007FFAE9E13615, C:\Python310\python310.dll(0x00007FFAE9DA0000) + 0x73615 byte(s), _PyObject_GC_Malloc() + 0x1055 byte(s) 0x00007FFAE9E17293, C:\Python310\python310.dll(0x00007FFAE9DA0000) + 0x77293 byte(s), _PyEval_EvalFrameDefault() + 0x19F3 byte(s) 0x00007FFAE9E13615, C:\Python310\python310.dll(0x00007FFAE9DA0000) + 0x73615 byte(s), _PyObject_GC_Malloc() + 0x1055 byte(s) 0x00007FFAE9E17293, C:\Python310\python310.dll(0x00007FFAE9DA0000) + 0x77293 byte(s), _PyEval_EvalFrameDefault() + 0x19F3 byte(s) 0x00007FFAE9E13615, C:\Python310\python310.dll(0x00007FFAE9DA0000) + 0x73615 byte(s), _PyObject_GC_Malloc() + 0x1055 byte(s) 0x00007FFAE9E17293, C:\Python310\python310.dll(0x00007FFAE9DA0000) + 0x77293 byte(s), _PyEval_EvalFrameDefault() + 0x19F3 byte(s) 0x00007FFAE9E13615, C:\Python310\python310.dll(0x00007FFAE9DA0000) + 0x73615 byte(s), _PyObject_GC_Malloc() + 0x1055 byte(s) 0x00007FFAE9E17293, C:\Python310\python310.dll(0x00007FFAE9DA0000) + 0x77293 byte(s), _PyEval_EvalFrameDefault() + 0x19F3 byte(s) 0x00007FFAE9E13615, C:\Python310\python310.dll(0x00007FFAE9DA0000) + 0x73615 byte(s), _PyObject_GC_Malloc() + 0x1055 byte(s) 0x00007FFAE9E17293, C:\Python310\python310.dll(0x00007FFAE9DA0000) + 0x77293 byte(s), _PyEval_EvalFrameDefault() + 0x19F3 byte(s) 0x00007FFAE9E13615, C:\Python310\python310.dll(0x00007FFAE9DA0000) + 0x73615 byte(s), _PyObject_GC_Malloc() + 0x1055 byte(s) 0x00007FFAE9E17293, C:\Python310\python310.dll(0x00007FFAE9DA0000) + 0x77293 byte(s), _PyEval_EvalFrameDefault() + 0x19F3 byte(s) 0x00007FFAE9E13615, C:\Python310\python310.dll(0x00007FFAE9DA0000) + 0x73615 byte(s), _PyObject_GC_Malloc() + 0x1055 byte(s) 0x00007FFAE9E17293, C:\Python310\python310.dll(0x00007FFAE9DA0000) + 0x77293 byte(s), _PyEval_EvalFrameDefault() + 0x19F3 byte(s) 0x00007FFAE9E13615, C:\Python310\python310.dll(0x00007FFAE9DA0000) + 0x73615 byte(s), _PyObject_GC_Malloc() + 0x1055 byte(s) 0x00007FFAE9E17293, C:\Python310\python310.dll(0x00007FFAE9DA0000) + 0x77293 byte(s), _PyEval_EvalFrameDefault() + 0x19F3 byte(s) 0x00007FFAE9E13615, C:\Python310\python310.dll(0x00007FFAE9DA0000) + 0x73615 byte(s), _PyObject_GC_Malloc() + 0x1055 byte(s) 0x00007FFAE9DFC00C, C:\Python310\python310.dll(0x00007FFAE9DA0000) + 0x5C00C byte(s), PyVectorcall_Call() + 0xB8 byte(s) 0x00007FFAE9DFBE87, C:\Python310\python310.dll(0x00007FFAE9DA0000) + 0x5BE87 byte(s), _PyObject_Call() + 0x143 byte(s) 0x00007FFAE9E1B5E7, C:\Python310\python310.dll(0x00007FFAE9DA0000) + 0x7B5E7 byte(s), _PyEval_EvalFrameDefault() + 0x5D47 byte(s) 0x00007FFAE9E18620, C:\Python310\python310.dll(0x00007FFAE9DA0000) + 0x78620 byte(s), _PyEval_EvalFrameDefault() + 0x2D80 byte(s) 0x00007FFAE9E149D7, C:\Python310\python310.dll(0x00007FFAE9DA0000) + 0x749D7 byte(s), _PyFunction_Vectorcall() + 0x87 byte(s) 0x00007FFAE9E16033, C:\Python310\python310.dll(0x00007FFAE9DA0000) + 0x76033 byte(s), _PyEval_EvalFrameDefault() + 0x793 byte(s) 0x00007FFAE9E149D7, C:\Python310\python310.dll(0x00007FFAE9DA0000) + 0x749D7 byte(s), _PyFunction_Vectorcall() + 0x87 byte(s) 0x00007FFAE9DFC00C, C:\Python310\python310.dll(0x00007FFAE9DA0000) + 0x5C00C byte(s), PyVectorcall_Call() + 0xB8 byte(s) 0x00007FFAE9DFBE87, C:\Python310\python310.dll(0x00007FFAE9DA0000) + 0x5BE87 byte(s), _PyObject_Call() + 0x143 byte(s) 0x00007FFAE9E1B5E7, C:\Python310\python310.dll(0x00007FFAE9DA0000) + 0x7B5E7 byte(s), _PyEval_EvalFrameDefault() + 0x5D47 byte(s) 0x00007FFAE9E149D7, C:\Python310\python310.dll(0x00007FFAE9DA0000) + 0x749D7 byte(s), _PyFunction_Vectorcall() + 0x87 byte(s) 0x00007FFAE9DFC00C, C:\Python310\python310.dll(0x00007FFAE9DA0000) + 0x5C00C byte(s), PyVectorcall_Call() + 0xB8 byte(s) 0x00007FFAE9DFBE87, C:\Python310\python310.dll(0x00007FFAE9DA0000) + 0x5BE87 byte(s), _PyObject_Call() + 0x143 byte(s) 0x00007FFAE9E1B5E7, C:\Python310\python310.dll(0x00007FFAE9DA0000) + 0x7B5E7 byte(s), _PyEval_EvalFrameDefault() + 0x5D47 byte(s) 0x00007FFAE9E149D7, C:\Python310\python310.dll(0x00007FFAE9DA0000) + 0x749D7 byte(s), _PyFunction_Vectorcall() + 0x87 byte(s) 0x00007FFAE9E13769, C:\Python310\python310.dll(0x00007FFAE9DA0000) + 0x73769 byte(s), _PyObject_GC_Malloc() + 0x11A9 byte(s) 0x00007FFAE9DFBFB0, C:\Python310\python310.dll(0x00007FFAE9DA0000) + 0x5BFB0 byte(s), PyVectorcall_Call() + 0x5C byte(s) 0x00007FFAE9DFBE87, C:\Python310\python310.dll(0x00007FFAE9DA0000) + 0x5BE87 byte(s), _PyObject_Call() + 0x143 byte(s) 0x00007FFAE9E1B5E7, C:\Python310\python310.dll(0x00007FFAE9DA0000) + 0x7B5E7 byte(s), _PyEval_EvalFrameDefault() + 0x5D47 byte(s) 0x00007FFAE9E16A94, C:\Python310\python310.dll(0x00007FFAE9DA0000) + 0x76A94 byte(s), _PyEval_EvalFrameDefault() + 0x11F4 byte(s) 0x00007FFAE9E16A94, C:\Python310\python310.dll(0x00007FFAE9DA0000) + 0x76A94 byte(s), _PyEval_EvalFrameDefault() + 0x11F4 byte(s) 0x00007FFAE9E149D7, C:\Python310\python310.dll(0x00007FFAE9DA0000) + 0x749D7 byte(s), _PyFunction_Vectorcall() + 0x87 byte(s) 0x00007FFAE9E13769, C:\Python310\python310.dll(0x00007FFAE9DA0000) + 0x73769 byte(s), _PyObject_GC_Malloc() + 0x11A9 byte(s) 0x00007FFAE9DFBFB0, C:\Python310\python310.dll(0x00007FFAE9DA0000) + 0x5BFB0 byte(s), PyVectorcall_Call() + 0x5C byte(s) 0x00007FFAE9DFBD93, C:\Python310\python310.dll(0x00007FFAE9DA0000) + 0x5BD93 byte(s), _PyObject_Call() + 0x4F byte(s) 0x00007FFAE9E88962, C:\Python310\python310.dll(0x00007FFAE9DA0000) + 0xE8962 byte(s), _PyRuntimeState_Fini() + 0x5F2 byte(s) 0x00007FFAE9E888DE, C:\Python310\python310.dll(0x00007FFAE9DA0000) + 0xE88DE byte(s), _PyRuntimeState_Fini() + 0x56E byte(s) 0x00007FFB5CFF37B0, C:\WINDOWS\System32\ucrtbase.dll(0x00007FFB5CFF0000) + 0x37B0 byte(s), wcsrchr() + 0x150 byte(s) 0x00007FFB5F8FE8D7, C:\WINDOWS\System32\KERNEL32.DLL(0x00007FFB5F8D0000) + 0x2E8D7 byte(s), BaseThreadInitThunk() + 0x17 byte(s) 0x00007FFB5FD514FC, C:\WINDOWS\SYSTEM32\ntdll.dll(0x00007FFB5FCA0000) + 0xB14FC byte(s), RtlUserThreadStart() + 0x2C byte(s) Any ideas?

TAGopolis

Win 11 24h2, 32gigs of ram, 3090 with 24g. Intel i7-9700k

Jay

Lora training tutorial? This or Wan would be both incredible :)

Thae Maestro

ERROR DURING GENERATION: CUDA error: an illegal memory access was encountered 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.

Thae Maestro

3090 windows 10, 3800x cpu, 32gb ram. Thanks in advance! :)

Leonid Wolf

Can you help me with the setup --share. I was unable to create a link for publication. Please check your internet connection or our status page.: https://status.gradio.app. Applying a preset and initializing information at startup... The last used preset "None" was not found or invalid, loaded "By default".

SECourses

Lora is bugged atm. I plan to look today hopefully to improve it

SECourses

Hi it is related to your pc. You see Gradio gave you some errors.

SECourses

I feel like something else inferred. Can you set 50 gb virtual ram, restart and try again without any other apps and let me know?

SECourses

Thanks will do hopefully but so many tasks :)

SECourses

Also set virtual ram to 50 gb at least

RC

For me it's not working, tried multiple LoRas, converting them, checking the box to convert and also without it. It just gets stuck at step 0.

cool1

I extracted frame pack v40.zip (as far as I know) in my frame pack folder, overwriting most of the other files (except start_app). I haven't run windows update (not sure i need to) but running it it still shows V21 at the top and I don't see a sliding window setting in the UI. Is it running the latest and is it understanding the timecode stuff? Or should I haven't run windows update.bat? edit: I've just run windows_update.bat and then restarted the app and now it's showing V41 at the top of the UI so I think that was my mistake not running windows update (though I'm not sure why unzipping v40.zip didn't increase the version number to 40).

Thae Maestro

I changed the preserved memory to 50GB. I think thats what you mean right? it generated 1sec of video and errors out generating the 2nd sec of video. this was with a fresh restart and i dont think anything else is running.: "* Running on local URL: http://127.0.0.1:7860 To create a public link, set `share=True` in `launch()`. Applying preset and initializing info on startup... Last used preset 'None' not found or invalid, loading 'Default'. Multi-line disabled: Passing full prompt to worker for potential timestamp parsing. Starting processing loop 1/1 Timestamped prompt format not detected or invalid, using the entire prompt as one. Unloaded DynamicSwap_LlamaModel as complete. Unloaded CLIPTextModel as complete. Unloaded SiglipVisionModel as complete. Unloaded AutoencoderKLHunyuanVideo as complete. Unloaded DynamicSwap_HunyuanVideoTransformer3DModelPacked as complete. Loaded CLIPTextModel to cuda:0 as complete. Found best resolution bucket 512 x 768 Saved start image to E:\AI\Framepack\FramePack\outputs\used_images\2025_04_25_19_14_30_447_start.png Unloaded CLIPTextModel as complete. Loaded AutoencoderKLHunyuanVideo to cuda:0 as complete. Unloaded AutoencoderKLHunyuanVideo as complete. Loaded SiglipVisionModel to cuda:0 as complete. latent_padding_size = 27, is_last_section = False, is_first_section = True Unloaded SiglipVisionModel as complete. Moving DynamicSwap_HunyuanVideoTransformer3DModelPacked to cuda:0 with preserved memory: 50 GB Teacache: 0.15 100%|██████████████████████████████████████████████████████████████████████████████████| 25/25 [04:10<00:00, 10.02s/it] Section completed in 250.44 seconds VAE decoding started (takes longer on higher resolution)... Using memory optimization: VAE offloading enabled Offloading DynamicSwap_HunyuanVideoTransformer3DModelPacked from cuda:0 to preserve memory: 50 GB Loaded AutoencoderKLHunyuanVideo to cuda:0 as complete. VAE decoding completed in 12.87 seconds (1.430 sec/frame) Unloaded AutoencoderKLHunyuanVideo as complete. Saved MP4 video to E:\AI\Framepack\FramePack\outputs\intermediate_videos\2025_04_25_19_14_30_447_9.mp4 Saving operations completed in 1.89 seconds Decoded. Current latent shape torch.Size([1, 16, 9, 96, 64]); pixel shape torch.Size([1, 3, 33, 768, 512]) latent_padding_size = 18, is_last_section = False, is_first_section = False Moving DynamicSwap_HunyuanVideoTransformer3DModelPacked to cuda:0 with preserved memory: 50 GB Teacache: 0.15 64%|████████████████████████████████████████████████████▍ | 16/25 [01:32<00:51, 5.77s/it] ================================================== ERROR DURING GENERATION: CUDA error: an illegal memory access was encountered 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. ================================================== Traceback (most recent call last): File "E:\AI\Framepack\FramePack\app.py", line 1223, in worker generated_latents = sample_hunyuan( File "E:\AI\Framepack\FramePack\venv\lib\site-packages\torch\utils\_contextlib.py", line 116, in decorate_context return func(*args, **kwargs) File "E:\AI\Framepack\FramePack\diffusers_helper\pipelines\k_diffusion_hunyuan.py", line 116, in sample_hunyuan results = sample_unipc(k_model, latents, sigmas, extra_args=sampler_kwargs, disable=False, callback=callback) File "E:\AI\Framepack\FramePack\diffusers_helper\k_diffusion\uni_pc_fm.py", line 141, in sample_unipc return FlowMatchUniPC(model, extra_args=extra_args, variant=variant).sample(noise, sigmas=sigmas, callback=callback, disable_pbar=disable) File "E:\AI\Framepack\FramePack\diffusers_helper\k_diffusion\uni_pc_fm.py", line 126, in sample x, model_x = self.update_fn(x, model_prev_list, t_prev_list, vec_t, order) File "E:\AI\Framepack\FramePack\diffusers_helper\k_diffusion\uni_pc_fm.py", line 99, in update_fn model_t = self.model_fn(x_t, t) File "E:\AI\Framepack\FramePack\diffusers_helper\k_diffusion\uni_pc_fm.py", line 23, in model_fn return self.model(x, t, **self.extra_args) File "E:\AI\Framepack\FramePack\diffusers_helper\k_diffusion\wrapper.py", line 37, in k_model pred_positive = transformer(hidden_states=hidden_states, timestep=timestep, return_dict=False, **extra_args['positive'])[0].float() File "E:\AI\Framepack\FramePack\venv\lib\site-packages\torch\nn\modules\module.py", line 1751, in _wrapped_call_impl return self._call_impl(*args, **kwargs) File "E:\AI\Framepack\FramePack\venv\lib\site-packages\torch\nn\modules\module.py", line 1762, in _call_impl return forward_call(*args, **kwargs) File "E:\AI\Framepack\FramePack\diffusers_helper\models\hunyuan_video_packed.py", line 983, in forward hidden_states, encoder_hidden_states = self.gradient_checkpointing_method( File "E:\AI\Framepack\FramePack\diffusers_helper\models\hunyuan_video_packed.py", line 832, in gradient_checkpointing_method result = block(*args) File "E:\AI\Framepack\FramePack\venv\lib\site-packages\torch\nn\modules\module.py", line 1751, in _wrapped_call_impl return self._call_impl(*args, **kwargs) File "E:\AI\Framepack\FramePack\venv\lib\site-packages\torch\nn\modules\module.py", line 1762, in _call_impl return forward_call(*args, **kwargs) File "E:\AI\Framepack\FramePack\diffusers_helper\models\hunyuan_video_packed.py", line 594, in forward hidden_states = gate * self.proj_out(hidden_states) File "E:\AI\Framepack\FramePack\venv\lib\site-packages\torch\nn\modules\module.py", line 1751, in _wrapped_call_impl return self._call_impl(*args, **kwargs) File "E:\AI\Framepack\FramePack\venv\lib\site-packages\torch\nn\modules\module.py", line 1762, in _call_impl return forward_call(*args, **kwargs) File "E:\AI\Framepack\FramePack\venv\lib\site-packages\torch\nn\modules\linear.py", line 125, in forward return F.linear(input, self.weight, self.bias) File "E:\AI\Framepack\FramePack\diffusers_helper\memory.py", line 26, in hacked_get_attr return torch.nn.Parameter(p.to(**kwargs), requires_grad=p.requires_grad) RuntimeError: CUDA error: an illegal memory access was encountered 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. ================================================== Error during memory cleanup after exception: CUDA error: an illegal memory access was encountered 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."

Male AI Art

I followed your tutorial and installed Python directly under the c drive, but I still get the error message. 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] 既存の接続はリモート ホストに強制的に切断されました。

Weronika

Edit this file and set set HF_HUB_ENABLE_HF_TRANSFER=0 if you get model download errors of HF_TRANSFER Hugging Face cache folders set to: C:\FramePack\models W0426 07:34:43.506348 33940 venv\Lib\site-packages\torch\distributed\elastic\multiprocessing\redirects.py:29] NOTE: Redirects are currently not supported in Windows or MacOs. Currently enabled native sdp backends: ['flash', 'math', 'mem_efficient', 'cudnn'] Xformers is not installed! Flash Attn is installed! Sage Attn is installed! Namespace(share=False) Free VRAM 5.03515625 GB High-VRAM Mode: False Downloading shards: 100%|██████████████████████████████████████████████████████████████| 4/4 [00:00<00:00, 1951.75it/s] Loading checkpoint shards: 100%|█████████████████████████████████████████████████████████| 4/4 [00:03<00:00, 1.25it/s] Loading checkpoint shards: 33%|███████████████████ | 1/3 [00:00<00:00, 8.97it/s]Press any key to continue . . .

cool1

Is there a reason why FramePack stops working overnight if the user's router is turned off and then later (when router starts turning on ?) it gives "error" overlays? what's it trying to do online that's needed? I'd edited the windows_start.bat (and kept the same one as your updates don't have the edit) to have "set GRADIO_ANALYTICS_ENABLED=False" before "python app.py" is called so shouldn't that stop gradio logging? If so why is it still failing to run overnight and giving the "error" overlays if the router is turned off overnight? What other logging is or connection is trying to happen? It only starts creating the video again once it has a connection. It shouldn't need one. Could it please be changed so it works without an internet connection (when it already has all the ai models downloaded)?

koon grizzly

i saw him reply and point to his own comment for this on an earlier version, at the moment it's not possible to use this offline :( doesn't even start without internet connection

cool1

Surely it should be possible if the whole project is Python based and all the Python code is there (on our local PCs) and available to edit. Surely the Python code that tries to unnecessarily connect and do whatever telemetry it's doing can be edited out if we know where it's doing it. Perplexity.ai said the "set GRADIO_ANALYTICS_ENABLED=False" should have done that but it seems not.

Brandon

Getting any error trying to install lastest Framepack when downloading the shards safetensors_rust.SafetensorError: Error while deserializing header: InvalidHeaderDeserialization

SECourses

related your internet. they are downloaded from hugging face. you can edit bat file set HF_HUB_ENABLE_HF_TRANSFER=0

SECourses

out of Ram set virtual ram to min 50 gb

SECourses

that means it was using shared vram. also during generation if you cause any vram usage increase by other apps it will move to shared vram usage and will be very slow

SECourses

just fixed lora with v43 please read top changelogs

SECourses

just fixed lora with v43 please read top changelogs

cool1

Thanks. So it's a bug in someone else's code section it seems, that seems to be trying to load or check a model online that it doesn't need to as at that point it should be working off-line and should already have that model locally. Hopefully he can fix it, thanks for pointing out where the problem is. So it was a different thing than the standard gradio UI trying to do telemetry in this case.

SECourses

just fixed lora loading vram problem please read chagelogs at the top

TAGopolis

Will give it a shot! Thank you.

puk

I installed the new version, but for some reason, the generation now takes twice as long. It used to take 50 seconds per frame, now it's 100 seconds. I'm using an RTX 4090.

SECourses

Doesn't make sense weird. Did you restart computer and tried? Did you check your vram usage? Let me do few test maybe deleting xformers can make difference than I will update

SECourses

thanks for reporting. update to v44 now fixed. the issue was teacache was disabled permanently with a mistake now fixed

KonekoFuta

Are there any plans to add an upscale option? I wonder if that would let us generate at a lower resolution for faster gens.

SECourses

100% working on it today with best available Ai model

puk

Thanks for the super-fast support. Now everything's working like before.

koon grizzly

Anyone having great success in trying to have the photo do what is mentioned in the prompt or timestamps when creating longer video ? i have various results, in my last attempt i asked grok to create these timestamp prompts (and have it pay attention for huyan model in framepack) I used an ai generated photo of a bear that is a dj and women are dancing in the background. Grok came up with this for a 30 second video for the photo. [1] Bear spins record, nodding to beat. Crowd sways under blue lights. [3] Bear tweaks mixer, lights turn purple, crowd gets hyped. [5] Woman starts dancing, subtle moves. [7] Spotlight hits woman, she dances fluidly. [10] Spotlight brightens, woman spins gracefully. [12] Bear scratches record, lights flash, crowd cheers. [15] Woman twirls under spotlight, captivating the crowd. [18] Close-up on woman, detailed dance under bright light. [20] Spotlight follows woman across dance floor, synced with beat. [22] Bear raises paw, lights pulse rapidly, crowd excited. [25] Woman joins crowd, raises hands in joy. [26] Beat drops, lights explode in colors, crowd jumps, woman celebrates. [29] Bear removes headphones, grins. Lights dim, music softens. In the end i got the bear to have a couple of different movements (which is better then older attempts where he did mostly the same or nothing at all) But the colors of the spotlights, the change of camera angle to the dancing woman (or basically any change at all in the background) was ignored and takes until the last 4-5 seconds Am i asking to much of this model or is there a trick to write better prompts ? Settings used : 640p resolution / latent windows size 8 / teacache : 0.15 / steps : 20 (result looked same to me) / cfg scale : 2 / cfg scale distilled : 10 / cfg rescale : 0 / gpu memory reserved : 6 (edit that was on v41, when my next attempt is ready i will install and test v44)

macmotu

And another awesome version! Your latest update with working LoRAs really made my day! Btw.: Do you ever sleep? ;-) P.S.: Will there be an option to select multiple LoRAs in the future?

Thae Maestro

I set the initial size and maxsize to 51200. then i restarted and tried again. heres the error after generating a video. :"Edit this file and set set HF_HUB_ENABLE_HF_TRANSFER=0 if you get model download errors of HF_TRANSFER Hugging Face cache folders set to: E:\AI\Framepack\FramePack\models W0426 12:44:15.744000 7980 venv\Lib\site-packages\torch\distributed\elastic\multiprocessing\redirects.py:29] NOTE: Redirects are currently not supported in Windows or MacOs. Currently enabled native sdp backends: ['flash', 'math', 'mem_efficient', 'cudnn'] Xformers is not installed! Flash Attn is installed! Sage Attn is installed! Namespace(share=False) Free VRAM 22.759765625 GB High-VRAM Mode: False Downloading shards: 100%|██████████████████████████████████████████████████████████████| 4/4 [00:00<00:00, 3635.37it/s] Loading checkpoint shards: 100%|█████████████████████████████████████████████████████████| 4/4 [00:00<00:00, 4.36it/s] Loading checkpoint shards: 100%|█████████████████████████████████████████████████████████| 3/3 [00:00<00:00, 8.29it/s] transformer.high_quality_fp32_output_for_inference = True Created batch outputs directory: E:\AI\Framepack\FramePack\outputs\batch_outputs Allowed Gradio paths: ['C:\\', 'D:\\', 'E:\\'] * Running on local URL: http://127.0.0.1:7860 To create a public link, set `share=True` in `launch()`. Applying preset and initializing info on startup... Last used preset 'None' not found or invalid, loading 'Default'. Multi-line disabled: Passing full prompt to worker for potential timestamp parsing. Starting processing loop 1/1 Timestamped prompt format not detected or invalid, using the entire prompt as one. Unloaded DynamicSwap_LlamaModel as complete. Unloaded CLIPTextModel as complete. Unloaded SiglipVisionModel as complete. Unloaded AutoencoderKLHunyuanVideo as complete. Unloaded DynamicSwap_HunyuanVideoTransformer3DModelPacked as complete. Loaded CLIPTextModel to cuda:0 as complete. Found best resolution bucket 640 x 640 Saved start image to E:\AI\Framepack\FramePack\outputs\used_images\2025_04_26_12_45_20_474_start.png Unloaded CLIPTextModel as complete. Loaded AutoencoderKLHunyuanVideo to cuda:0 as complete. Unloaded AutoencoderKLHunyuanVideo as complete. Loaded SiglipVisionModel to cuda:0 as complete. latent_padding_size = 27, is_last_section = False, is_first_section = True Unloaded SiglipVisionModel as complete. Moving DynamicSwap_HunyuanVideoTransformer3DModelPacked to cuda:0 with preserved memory: 8 GB Teacache: 0.15 100%|██████████████████████████████████████████████████████████████████████████████████| 25/25 [02:33<00:00, 6.14s/it] Section completed in 153.54 seconds VAE decoding started (takes longer on higher resolution)... Using memory optimization: VAE offloading enabled Offloading DynamicSwap_HunyuanVideoTransformer3DModelPacked from cuda:0 to preserve memory: 8 GB Loaded AutoencoderKLHunyuanVideo to cuda:0 as complete. VAE decoding completed in 14.55 seconds (1.617 sec/frame) Unloaded AutoencoderKLHunyuanVideo as complete. Saved MP4 video to E:\AI\Framepack\FramePack\outputs\intermediate_videos\2025_04_26_12_45_20_474_9.mp4 Saving operations completed in 1.28 seconds Decoded. Current latent shape torch.Size([1, 16, 9, 80, 80]); pixel shape torch.Size([1, 3, 33, 640, 640]) latent_padding_size = 18, is_last_section = False, is_first_section = False Moving DynamicSwap_HunyuanVideoTransformer3DModelPacked to cuda:0 with preserved memory: 8 GB Teacache: 0.15 0%| | 0/25 [00:10

Dan

The man is insanely productive!

SECourses

thanks a lot. multiple lora i may add in future.

SECourses

the model may not be that much capable after all

koon grizzly

it does a great job creating smaller video, 5 second videos to zoom in/out or have other movements and use last frame as 1st frame give nice results.

koon grizzly

Maybe we can have an option that will automate using the last frame of created video and start new video with same settings with that frame ? like that i can go to bed and wake up with lot's of videos that follow up on each other :-)

koon grizzly

Yes same for me, i think my longest is about 70 seconds in 11 hours or longer (on 4060 with 8 gig) For testing purpose i will put teacache to 0.5 and sampling to 15 instead of 25 and resolution at 480. It will go double as fast (between 2.5 & 3.5 min per second) but there will be lots of blurry parts because of the teacache at 0.5 so its only useful to quickly test a couple of seconds

SECourses

nothing wrong. 30 seconds is a huge time for your GPU at that resolution. you can reduce resolution to speed up or increase teacache or reduce number of steps

SECourses

yes it is good to test at low res first

SECourses

nice no issues working. i recommend make the preserve memory = 12 gb

SECourses

i dont see failed here. if your app terminating that means you are out of RAM. if you cant solve set virtual RAM to 50 GB and make a fresh install in a new folder

Male AI Art

Is it conceivable that security software may be involved? I've never asked the question on github, but I'll do my best to try as best I can.

Steve

Cool concept, but I also agree with some other comments that long videos dont really work. Dont know if its the app or some settings. The motion stops and slows to 0 for videos longer than 15 seconds. I tried no lora, no teacache. Its very evident for long videos over 30 seconds. The video starts as a very noisy image, then no motion until about the last 20 seconds, and is still slow motion. The motion then speeds up only at the final 10 seconds. I see on the original github that long videos are possible, must be some setti gs that are broken. Update: After some experimentation, increasing the latent window size seems to help make longer videos more consistent, of course requires more memory and time but works

koon grizzly

Thanks for the tip, i tested it and video to video test worked for me but photo to video 480p just keeps working (more then 4000 seconds meanwhile) is that normal for 4 gigabyte option in rtx4060 (i am already looking around for a cheap used 3080 lol) If so i Framepack is a much faster solution for me (and i am happy with the results it creates when using last frame to start next short video : here is my latest test results : https://youtube.com/shorts/kdCck29biwk?si=5pn3lZxwuGxnszLD

Thae Maestro

:( it was still broken but i figured it out. my 3090 vram was overheating and it was crashing.

SECourses

i recommend 2 hand used rtx 3090 if you can find. high vram makes huge diff

Leonid Wolf

I have tested more than 15 videos of 5-10 seconds on different styles. And I realized one thing. That the Frame Pack works better on cartoon styles. In realism, artifacts constantly appear that cannot be removed only by postproduction to remove watermarks. When a video is generated at 10 seconds, the character starts slowing down. I think it's okay. Since such models that are publicly available do not produce the result that you expect. I tried to do the generation in the comfortable ui with the LTX and hunyuan models, the results are the same.

SECourses

thanks a lot for the observations. i think if framepack was based on wan 2.1 that would be really different level

SECourses

100% your anti-virus could be causing this

koon grizzly

yeah, but used 3090 is still 700-1000 euro, that is way over my budget for having a little bit of fun :-) i bought a 3080 last night for 300, will be sent to me this week. According to my grok chat bot it was the best value for money to get compared to my 4060 (The RTX 3080 10GB will be approximately 50–60% faster than the RTX 4060 8GB for creating a video from a photo in Framepack, due to its higher CUDA core count (8704 vs. 3072), more Tensor cores (272 vs. 96), and greater VRAM/bandwidth (10GB vs. 8GB). I have to figure out, but maybe i can use both the 4060 and the 3080?

SECourses

in some models and apps you can load different parts on each gpu but it requires some more technical knowledge and editing. 3080 10 gb should be faster than 8gb 4060 true

koon grizzly

i gave up on the wan2.1, photo to video, the model for photo to video takes hours for a couple of seconds and video to video changed everything to much (faces from people change, my realistic dj bears become cartoonish, i tested all settings, it don't want to stay consistent) Oh yes, is there an uncensored text to photo model that can do same quality like grok3? I have been testing Sana v17 which works great and fast but none of the included image styles or options (even with sampling steps at 100 etc) it cannot come close to the quality i get from the grok3 online model when it comes to creating a bear deejay.

koon grizzly

Thanks for the tip on the latent window, i will do some testing again with longer videos

TriVectorX

Quick question, I trained my face in Hunyuan video as Lora, it looks almost identical when using ComfyUI + Hunyuan workflow; However, when I use the same lora in this frampack, it doesn't fully take my face lora, and the character looks like a mixture of my face with someone else. I tried to adjust the weight from 1.0 to 2.0, not really work as expected. What could be the root cause here?

Nicolas Ruarte

Hi trully great work with this installation is the best FramePack distribution out there, thank you. I have a suggestion for the next version I've been testing this distribution recently launched by red_polo on github Frame Pack Loop, it works great for loops up to 30 seconds and uses the same instalation as the OG FramePack, could you see to add it to your instalation? https://github.com/red-polo/FramePackLoop

Walker4k

H100 at MC gives this error on FRamepack and Wan - File "/home/Ubuntu/Downloads/FramePack_v44_v2/FramePack/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 wonder if sage attention not supporting H100. can you uninstall and try? or you can pick A100 or RTX A6000 ADA. i am gonna work on this now

SECourses

what is it doing can you explain me in details please

SECourses

yes not fully compatible sadly :( probably you need to train framepack lora which kohya added experimentally support atm

SECourses

i found the mistake. working on fix sorry for this. it was my pre-compile error

Walker4k

Both FP and Wan fixed? Had the same error with both

Hipno

Omg this is actually insane!! Does this also work well with simple 2d cartoon styles aswell?

koon grizzly

I first tested latent window size at 10 and didn't see much improvement. Before i left for work i started another 60 second video, this time with latent window size at 30 (other settings 30fps 480p teacache 0.15 and steps at 20) Oh and there is a [ ] timestamp prompt every 3 seconds because giving timestamps does something to get different movements compared to giving just 1 prompt for the whole video. I just got back home, 10 hours and 33 minutes later i got a new 60 seconds video and i am quite happy with the result. 1. The video runs faster then normal, i had to slow it down about 33% in my vlc player to have normal dancing people in the background so it's actually a 90 second video. 2. Nothing is Not moving, the background does mostly the same but that's ok, I think the prompts helped to get at least 4 different movements (that repeat about 10 seconds) and i was happy to notice the spotlights also move direction at some point (i am gonna pay more attention to that in the timestamps next time) 3. My testdj bear is giving me a good result, (almost as good as my results with creating short 3-5 seconds videos where i use last frame to start the new video to get one long video with lot's of different moves) -> he does things on his mixer, even grabs his headphones at one point (and the usual 'pretending' to do something while hovering above the dj gear and the shaking of the hands/paws is also there but at least it's no longer 50 seconds only doing that. a video says more then all these words so check for yourself 1 is from a week ago and was my first test to get a long video https://www.youtube.com/watch?v=wgPPrcT0j0k 2 is the new video with the latent window setting at 30 (i uploaded the unedited version so it is still fast movement, the music i chose allows this fast dancing lol) https://www.youtube.com/shorts/OWgYdAu7L7A The difference compared to my test a week ago is HUGE so Thanks again for the tip, for my fun project this is what i wanted, now i don't have to spend all day anymore to create all these little videos that start woth last frame from the previous one

SECourses

i think it works even better with that :D

SECourses

yes they are both fixed. i have updated the remote url. i tested FP it works. so wan should work too. sorry for this

Hipno

That's so great to hear!! Honestly you're doing so much for the community and I respect you so much for that!! 👍🏽👍🏽

koon grizzly

Maybe 20 will do also, i am now using latent window set to 25 with the photo from a week ago (that gave the slow motion video) to make sure it was not the photo that caused the lack of movement I am happy to see that last 20 seconds are done and the movement is really good and every second keeps going at the same speed etc. 25 give a little bit of extra speed, probably 25% so 20 should be normal speed Edit 21 was my next test last, 20 seconds done, last 10 seconds it starts speeding up, last 5 seconds it messed up. so 25 gave me best results (and also quite fast compared to 30)

Walker4k

Both now work beautifully, many thanks. Fingers crossed Wan gets onto Framepack soon

SECourses

still working sorry for delay. ty

koon grizzly

I have some tips for faster generation. I went from 3, 4, 5 hours to a little more then an hour for creating 30 seconds of footage 1. check my settings below and test them to find out if it works for you too Most important is latent window size at 25 and Fps also at 25, this will create 4 second (3.87) parts instead of 1 second parts which seems to go much faster because it doesn't have to stop after every second created. I used to have around 3-6 minutes per second, now i get 4 in about 6 minutes. (4060 gpu) 2 Use Smal size jpg files !! i noticed using smaller size photo uses less time between the generation of the video parts so for testing i created 2x a 30 second video with 2x exactly the same photo, settings and prompts but.. one was 100kb and the other had reduced quality to make it 50kb. This reduced the generation time even more to 48 minutes. Edit : Not sure if it was the photo or the seed but the smaller jpg seems to have faster movement then the 100kb version (i watched youtube & did nothing else during both generations) test 100kb _________ Negative Prompt: Seed: 1158344243 TeaCache: Enabled (Threshold: 0.15) Video Length (seconds): 30 FPS: 25 Latent Window Size: 25 Steps: 15* CFG Scale: 1 Distilled CFG Scale: 10 Guidance Rescale: 0 Resolution: 480 Generation Time: 1 hour 7 min 31 seconds Total Seconds: 4051 seconds Start Frame Provided: True End Frame Provided: False Timestamped Prompts Used: True test 50kb ________ Negative Prompt: Seed: 688717800 TeaCache: Enabled (Threshold: 0.15) Video Length (seconds): 30 FPS: 25 Latent Window Size: 25 Steps: 15* CFG Scale: 1 Distilled CFG Scale: 10 Guidance Rescale: 0 Resolution: 480 Generation Time: 47 min 48 seconds Total Seconds: 2868 seconds Start Frame Provided: True End Frame Provided: False Timestamped Prompts Used: True * I have steps at 15 but using 20 or even 25 for better details should not double the time needed.

koon grizzly

2 minute video with teacache at 0.1 & 20 steps : 6.5 hours, i have 1 minute videos that took 11 hours. (jpg used had quality down to 25% of original to make it only 75kb) Seed: 214271538 TeaCache: Enabled (Threshold: 0.1) Video Length (seconds): 120 FPS: 25 Latent Window Size: 25 Steps: 20 CFG Scale: 1 Distilled CFG Scale: 10 Guidance Rescale: 0 Resolution: 480 Generation Time: 6 hours 28 min 0 seconds Total Seconds: 23280 seconds Start Frame Provided: True

koon grizzly

Next step : find prompts that work, if anyone has tips please tell me Check the 2 minute video, (it has the timestamped prompts in the description), only thing that happened that was special was the bear is blowing bubbles at some point buit other then that it's mostly the same https://www.youtube.com/watch?v=ieAK0Hi8hJY

koon grizzly

All credits go to you for providing us with these settings. I do hope you can find a way to give us batch/automated use of last frame. It would be perfect if it could use also use a new text prompt when it takes the last frame to start new one because it is easier to create short video with only one prompt to get good results for action in the video. Like that we could create full movies while we do other things

SECourses

framepack already has longer video generation why do you need get last frame feature? wan 2.1 app has it

SECourses

and we already support time based prompts. so you can use different prompts for different parts of the video

koon grizzly

If you check my 2 minute video you will notice it is 2 minutes mostly doing the same thing, prompts are barely followed or not at all while i had one for every 4 seconds, i keep changing the way the prompts are written while hoping for a change and this time i even used cartoonish creatures instead of 'real' people but the results stay the same. I do get pretty good results when using only 1 prompt in a 5 or 10 second video so my wish is to have many 5-10 second videos that compile into 1 at the end without the need to stay at the computer all day long. I tested wan for 2 days, the photo to video model is way to heavy for my computer, takes hours for a couple of seconds and results look horrible compared to this model. The video to video model goes faster but it can't stay consistent, it doesn't need a second to start morphing pople into someone else, or in my testing it turns the detailed bear into a cartoonish bear

SECourses

the problem with extending is that the last frame quality is not being great so the quality constantly degrades. i can implement but requiring a lot of work.

Bjoern

Do I need any "special type" of LoRas or can I use the WAN LoRas? I tried one (with and without converting), but I always get this error: Loading new LoRA structure: squish_18_default from G:\FramePack_v23\FramePack\loras\squish_18_default.safetensors ERROR loading LoRA structure for squish_18_default: list index out of range Traceback (most recent call last): File "G:\FramePack_v23\FramePack\app.py", line 1563, in manage_lora_structure load_lora(transformer, lora_dir, lora_filename) # Assume modification in-place on CPU model File "G:\FramePack_v23\FramePack\diffusers_helper\load_lora.py", line 36, in load_lora transformer.load_lora_adapter(state_dict, network_alphas=None, adapter_name=weight_name.split(".")[0]) File "G:\FramePack_v23\FramePack\venv\lib\site-packages\diffusers\loaders\peft.py", line 264, in load_lora_adapter lora_config_kwargs = get_peft_kwargs(rank, network_alpha_dict=network_alphas, peft_state_dict=state_dict) File "G:\FramePack_v23\FramePack\venv\lib\site-packages\diffusers\utils\peft_utils.py", line 153, in get_peft_kwargs r = lora_alpha = list(rank_dict.values())[0] IndexError: list index out of range LoRA Management Error: Failed to load LoRA structure 'squish_18_default'.

SECourses

yes you need hunyuan loras. squish_18_default is for Wan 2.1

koon grizzly

ok, not needed that hard :-) And you are correct, they degrade after a couple of runs, especially when the face is not looking at camera at the end frame. To avoid this i use the original frame as last frame sometimes. i just got that 3080 installed and doing some first test runs (without changing settings or reinstalling something) - Test 1 Seed: 1098302674 TeaCache: Enabled (Threshold: 0.1) Video Length (seconds): 30 FPS: 25 Latent Window Size: 25 Steps: 20 CFG Scale: 1 Distilled CFG Scale: 10 Guidance Rescale: 0 Resolution: 480 Generation Time: 40 min 26 seconds (->10 min faster then 4060 for same settings, except i put vram to 9 instead of 7) Total Seconds: 2426 seconds - test 2 Seed: 1308756423 TeaCache: Enabled (Threshold: 0.1) Video Length (seconds): 30 FPS: 24 Latent Window Size: 30 Steps: 20 CFG Scale: 1 Distilled CFG Scale: 10 Guidance Rescale: 0 Resolution: 480 Generation Time: 35 min 35 seconds Total Seconds: 2135 seconds - Test 3 Seed: 1181341713 TeaCache: Enabled (Threshold: 0.1) Video Length (seconds): 30 FPS: 20 Latent Window Size: 31 Steps: 20 CFG Scale: 1 Distilled CFG Scale: 10 Guidance Rescale: 0 Resolution: 480 Generation Time: 30 min 14 seconds Total Seconds: 1814 seconds In every test i lowered the text prompts to fit the 4, 5 or 6 second parts which will also have lowered the total time needed for creations Test 3 has slower video play, test 2 is still ok, (i put on 1.1x playback speed) but test 3 needed 1.5x playback speed for me to like the movements

koon grizzly

try this setting (and don't use large & high quality photo), it prefers 50-100kb jpg more then 1 megabyte photo : FPS: 25 Latent Window Size: 25 Steps: 20 (25 looks better, takes more time) It will create 4 second parts instead of 1 second parts in about same time you can lower fps + raise latent window size to get 5 or 6 seconds per part but finished video becomes slower, need more testing but i think quality also goes down at 6 second parts. 4 seconds parts look good.

Lawren

This really needs a way to drag and drop or upload metadata to be reused instantly. Also, There's no metadata for intermediate videos it seems. Only final outputs. It would be ideal to be able to drop your intermediate video file into a tab and have it read the metadata? Frame Pack is absolutely phenomenal for what you can do in comparison of simplicity.

koon grizzly

I am making progress with creating/searching for prompts that work in longer videos. I think giving a lot of info doesn't work, it has to be really short and simple (at least for me) to have some succes. Below are some prompts it followed in my latest 2 minute test They jump up and down. They drink a glass of champagne. (champagne showed up) They are dancing They throw confetti. (confetti showed up) They wave to the crowd. They blow party horns. (party horn showed up) Pillow fight with cushions. (pillows showed up) They toss beach balls. (beachball showed up) They high-five They use glow sticks (something started glowing) They shake maracas. (Maracas showed up) They hug They blow bubble wands (bubbles showed up) Seed: 1176021196 Video Length (seconds): 120 FPS: 25 Latent Window Size: 25 Steps: 25 Generation Time: 3 hours 48 min 22 seconds

SECourses

metadata loading can be made from saved txt files. i will check that. intermediate videos metadata is same there isnt any difference that is why we dont have

Lawren

Yeah, the final output has an option to save the metadata, which I always do. I'm looking back at some of the intermediate_videos I stopped mid process, and there's no way to find out what settings I used for them. At least that's what it seems. Is this correct? Intermediate videos have no metadata attached to output.

koon grizzly

Another thing i found out : i noticed the model used my full 32 gig ddr (3200mhz) memory. (used) ddr4 is cheap so i replaced them with 64 (3600mhz) gig. Now my memory doesn't get full anymore (the model adds up to 50 gigabyte to the ddr memory) Did a quick test of 30 seconds with a (lowered to 50kb quality) jpg file. It started faster with the sampling of the steps of the first video part, sampling steps seems same speed. (This test also has 26fps and 25 steps) Seed: 1142625593 TeaCache: Enabled (Threshold: 0.1) Video Length (seconds): 30 FPS: 26 Latent Window Size: 25 Steps: 25 CFG Scale: 1 Distilled CFG Scale: 10 Guidance Rescale: 0 Resolution: 480 Generation Time: 39 min 8 seconds Total Seconds: 2348 seconds

SECourses

thanks. i saw and will look into that hopefully soon

SECourses

yes intermediate videos dont have. i see your point. ok i can add an option for that too. please remind me tomorrow both things

koon grizzly

Could the latent window size be implemented in the batch processing option ? last night i tested batch processing, woke up to a folder with files with videos with bad/slow/no movement. The latent window size really is important to have. EDIT, nevermind, it was because i had no (timestamped) prompts, i now notice it does use the settings from latent window size when creating the video and the first video i create with a prompt works normal Edit 2 wow they released a n update for Framepack ? woohoo Edit 3 : i quickly tested it and (without prompt) i see the same type of video, the blurry people in the background go to fast and people in the shot keep tapping the thing that was most close to their hands (a problem i see a lot) However, next 5 second test i gave it a prompt that people had to dance and yes i now have 5 seconds of dancing people pretty normally (also the blurry people in the background dance at the same speed)

SECourses

thanks. i am aware give me some time :D

koon grizzly

thanks again, without your modded version this video clip would have been impossible to create (and take 3x longer to create the video parts, i already notice the difference in the F1 version, took 15 minutes for 5 seconds lol, i got used to 2-3 minutes for 4 seconds)

SECourses

ye i will add it asap hopefully

koon grizzly

finished a 30 second test with the new model, i asked for the bear to drink a glass of champagne, fireworks come down from the sky and everyone is dancing (gracefully with clear movements) It did a really good job. The bear started drinking champagne at the start and he kept drinking for 30 seconds, same for the fireworks, it kept coming for 30 seconds. The dancing was not like i wanted but they were moving and it was not slow down (or fast speed in the background). Really great update

koon grizzly

new model compared to old model : when i ask to turn around the camera will move around my subject but in the old model it always came back to the original position, same for zoom in, it zoomed in a little and came back. In the new model when it turns ariund a subject or zooms in, it goes way further and it doesn't come back to it's original position. This opens great options for last frame usage :-)

SECourses

100% . just implemented F1 and working

cool1

Thanks for the update. So the F1 option should be faster but may or may not drift or lose quality compared to the other one it seems like. I assume this version could use less intermediate files or less total file size for them. Does it still stop working if there's no internet connection and if so are we still waiting for another developer to fix that or could it be some piece of code we could edit that could fix that?

puk

Hello Furkan, thank you for the new update. I have now been working with it for three hours and have realised that the generation with the F1 model on my 5090 takes much longer than before. I need almost 50 minutes for a 20 second video with 640 resolution. With the old model it takes just under 20 minutes. The results are very good with F1, but it's definitely too slow compared to the old model. The first 4 to 5 seconds at 0.20 TeaCache Threshold and 25 steps seem to render very quickly, but from second 8 the VEA encoding gets slower and slower

SECourses

very likely using shared vram. increase GPU Inference Preserved Memory (GB) by 4 gb and try again

SECourses

i am not sure yet didnt have chance to test. so try both

puk

THX for the quick answer. I've tried it all. The problem is with both models. A new installation did not fix the error. Fortunately, I didn't delete the previous version. This runs perfectly.

koon grizzly

Thanks for leaving the old version inside and available 'with a click of a button' Now we can still create transitions between two images. The new model is more responsive, haven't tested very long videos and timestamps yet but so far looks great and works same speed like old version for me

cool1

Thanks for the update. on the FramePack F1 option does the end image not work at all? I asked it to create an 8 second video and put the same image for the start and end frame (so I could create a continuous .mp4 & gif). I didn't notice the part where it said "optional (original model primarily)." . So what does it mean by "primarily". Does it mean it sometimes works on the F1 but not always or that it's not really meant for that? When it generated the video I asked for it to create an 8 second one of it created a 9 sec 634 ms video where the the first frame was quite a lot different tot the last frame - so it can't loop. So if we want repeatable videos (like .gifs) I assume it's best to use the main FramePack not F1 (even if that means it's slower). Though I haven't done a time comparison.

koon grizzly

I just saw something on yt, it has Queuing for creating videos . (edit Framepack studio, it's on official framepack website) So you take photo one, add prompt and start. Then you can take photo 2 add prompt and add it to the list. Add as much as you want but all with their own prompt. So you can create videos with different prompts that will auto start creating. I used batch before in this version, but it followed one prompt for every photo Can i do something like this with batch convert in Gradio?

Zachary Garner (edited)

Comment edits

2026-09-07 10:00 This takes forever to make a video on runpod Lol. I hope more updates to this makes it faster. Good work tho, good updates in there
2025-05-07 15:08

SECourses

i am not sure i didnt test intensively yet

SECourses

queue easy. your gradio interface on a new tab and gradio will auto queue

koon grizzly

so i open a new gradio window for every video i want to make and it will start next when one is done ?

Max Gonzales

Hi! It seems that framepack changes video (camera movement and coherence) on same seed when rendering at a higher resolution Does anyone know why this is happening? See examples below https://www.dropbox.com/scl/fi/2n6bkw1s65eh4wfeez51o/2025_05_08_11_37_26_820_37frames.mp4?rlkey=wp71lizqzdky13r6qc87vo06x&dl=0 & https://www.dropbox.com/scl/fi/iekdrggi0ozgf1e6d24qn/2025_05_08_11_41_08_726_37frames.mp4?rlkey=xsikag74pl9ekhp18ttzor78u&dl=0 Both are using the same seed - 1969421642

Lawren

Hey, Furkan, Just a quick reminder if you believe that would be a reasonable, and quick update. I would find it very helpful to have a drag and drop/upload metadata reader. I don't care what Reddit says, I think your work deserves appreciation. Thanks,

SECourses

high resolution looks much better. true different res behaving differently sadly

puk

Hello Furkan, i have reinstalled version 46 several times now and even deleted the .chache folder each time, the problem still persists. in version 45 i need 347.44s (5.79 min) for 10sec at 640. in version 46 with the old model 481. 52s (8.03 min) F1 Model 654.24s (10.90 min) with identical settings and seed. from 8 seconds the VAE encoding becomes slower and slower. if I generate 30 seconds it takes over 45 minutes at the end up to 3 minutes VAE encoding per second. With version 45 only 20 minutes. in version 45 xformers is not loaded, in version 46 xformers is loaded. can this be the reason? or are there other .cache folders that I should delete so that everything is freshly loaded from the servers during installation? I have not reloaded the models again.

SECourses

not it cant be reason. you can return older v45 simply with : open a cmd on FramePack folder and type git checkout ff23233b815a6abc6dbc0455bc0b14de207b9643 i will test speed for you now but i dont think there will be any difference

SECourses

yes i just tested. there isnt meaningful difference 621 seconds vs 557 seconds

umut yilmaz

8 may update is v.47 but latest installer zip is still v.46 😉

stphnvdb321

Thanks for V48. I've tested text to video (without an image). Like in FP Studio the picture quality is changing and getting worse with each new generated segment. :(

koon grizzly

when i try 16:9 format (640*360, 568*320 or 854*480) it does not want to start, it works when i change to 640*480. Maybe we need to remove the resolution output selector to be able to change the width and height because i had no problem selecting 640*360 or 568*320 in framepack studio EDIT : meanwhile restarted everything and it started with 640*360 Edit2 if other people get this problem, you need to save your old saved settings to a new save file so it has height and width included, it will work after that

koon grizzly

i ran update.bat file and now v48 :-)

stphnvdb321

This isn't the right place to ask, but can you make an installer for LTXV 13B 0.97?

DanO..

That's a problem from the very beginning of AI image generation. Same seed with different resolution gives totally different image. It's to be expected with video too.

SECourses

ah nice point. so it was older preset error

SECourses

yes this is the model limit i think

SECourses

yes zip file didnt need an update just run windows update :)

stphnvdb321

This would be great. I don't want to learn ComfyUI. I heard that this model should be 30 x faster.

SECourses

ye i dont like learning it either. i have been working on a workflow today and it is nightmare

koon grizzly

i have been watching tutorials today to get LTX running with all the extras in comfyui, i gave up halfway and went back to framepack :-) Also i wonder if it will run on 10 gigabyte GPU since the youtubers mean 4090 and 5090 when they talk about 'consumer grade gpu' lol Also, i read in some comments ltx is way more censored then framepack.

koon grizzly

I forgot to metion but when i did update bat file from v46 i got this error (but everything works fine i think) : Looking in indexes: https://download.pytorch.org/whl/nightly/cu128 ERROR: Could not find a version that satisfies the requirement torch==2.7.0.dev20250311 (from versions: 2.8.0.dev20250313+cu128, 2.8.0.dev20250314+cu128, 2.8.0.dev20250315+cu128, 2.8.0.dev20250316+cu128, 2.8.0.dev20250317+cu128, 2.8.0.dev20250318+cu128, 2.8.0.dev20250319+cu128, 2.8.0.dev20250320+cu128, 2.8.0.dev20250321+cu128, 2.8.0.dev20250322+cu128, 2.8.0.dev20250323+cu128, 2.8.0.dev20250324+cu128, 2.8.0.dev20250325+cu128, 2.8.0.dev20250326+cu128, 2.8.0.dev20250327+cu128, 2.8.0.dev20250401+cu128, 2.8.0.dev20250402+cu128, 2.8.0.dev20250403+cu128, 2.8.0.dev20250404+cu128, 2.8.0.dev20250405+cu128, 2.8.0.dev20250406+cu128, 2.8.0.dev20250407+cu128, 2.8.0.dev20250408+cu128, 2.8.0.dev20250409+cu128, 2.8.0.dev20250410+cu128, 2.8.0.dev20250411+cu128, 2.8.0.dev20250412+cu128, 2.8.0.dev20250413+cu128, 2.8.0.dev20250414+cu128, 2.8.0.dev20250415+cu128, 2.8.0.dev20250416+cu128, 2.8.0.dev20250417+cu128, 2.8.0.dev20250418+cu128, 2.8.0.dev20250419+cu128, 2.8.0.dev20250420+cu128, 2.8.0.dev20250421+cu128, 2.8.0.dev20250422+cu128, 2.8.0.dev20250423+cu128, 2.8.0.dev20250424+cu128, 2.8.0.dev20250425+cu128, 2.8.0.dev20250426+cu128, 2.8.0.dev20250427+cu128, 2.8.0.dev20250428+cu128, 2.8.0.dev20250429+cu128, 2.8.0.dev20250430+cu128, 2.8.0.dev20250501+cu128, 2.8.0.dev20250502+cu128, 2.8.0.dev20250503+cu128, 2.8.0.dev20250504+cu128, 2.8.0.dev20250505+cu128, 2.8.0.dev20250506+cu128, 2.8.0.dev20250507+cu128, 2.8.0.dev20250508+cu128, 2.8.0.dev20250509+cu128, 2.8.0.dev20250510+cu128) ERROR: No matching distribution found for torch==2.7.0.dev20250311

SECourses

this is older zip file. please use latest and delete older venv

Taiga

Fresh install of v46. Windows_Start_APP.bat. No interface comes up, just this: Hugging Face cache folders set to: D:\FramePack_v46\FramePack\models W0511 23:28:51.781000 51900 venv\Lib\site-packages\torch\distributed\elastic\multiprocessing\redirects.py:29] NOTE: Redirects are currently not supported in Windows or MacOs. Currently enabled native sdp backends: ['flash', 'math', 'mem_efficient', 'cudnn'] Xformers is installed! Flash Attn is installed! Sage Attn is installed! Namespace(share=False) Free VRAM 20.9609375 GB High-VRAM Mode: False Downloading shards: 100%|███████████████████████████████████████████████████| 4/4 [00:00

adpunk

Hi. Appreciate all the great FramePack work! Is there a link to v48? The most recent post links to v46 only.

koon grizzly

it is v46 zip file, there is no newer :-) i redownloaded from both links on the page and retry a couple of times. Error stays. But everything works so i guess it's no problem

koon grizzly

i have something similar today, - test 2 minute video in framepack F. (batch) - One prompt It started with 1 minute and a couple of seconds on average for 1 second of footage, it stayed like that for the first 20 seconds after that it started going slower and slower and by the time i got to the end it needed more then 10 minutes for 1 second. it's like between the parts (after it finishes the 30 steps) everything took a break.

Jonson Jonson

Thanks. Any way to train a LORA using 5000 series GPU? I have found Kohya and Musubi, but neither have 5000 series support yet.

SECourses

it is installing v48 but i just updated zip file to prevent more confusion :D

Jonson Jonson

Thanks. Am I right in thinking you can't train a Hunyuan LORA/finetune with that though? I am aiming to use a LORA on FramePack.

SECourses

you can train a lora for framepack. kohya made implementation but unofficial so we cant say works perfect

Taiga

Sorry, you helped with my last issue. Thought you could help with this: Total worker time for 1 generation(s): 700.33 seconds (11.67 minutes) Unloaded DynamicSwap_LlamaModel as complete. Unloaded CLIPTextModel as complete. Unloaded SiglipVisionModel as complete. Unloaded AutoencoderKLHunyuanVideo as complete. Unloaded DynamicSwap_HunyuanVideoTransformer3DModelPacked as complete. ERROR:asyncio:Exception in callback _ProactorBasePipeTransport._call_connection_lost(None) handle: Traceback (most recent call last): File "C:\Users\AI\AppData\Local\Programs\Python\Python310\lib\asyncio\events.py", line 80, in _run self._context.run(self._callback, *self._args) File "C:\Users\AI\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

Jonson Jonson

Using a SDXL base model, or another one?

SECourses

yes gradio error. vpn or antivirus may cause. still the work should be done and saved in outputs folder

SECourses

you need to use hunyuan i believe but i didnt test

Jonson Jonson

I tried the Hunyuan one and it wouldn't work. Thanks for replying.

patrikeon

Hi! Is there a way to start the gui (Windows_Start_APP.bat) without being online. I’m getting errors saying for example ConnectionErrror. A common error and an easy fix? Thank you Furkan! /P

Ec Jep

excellent work here! do you plan to create a training video on how to train wan2.1 with diffusion pipe or Musubi Tuner? I used your previous koyha training videos and extensively and hope to see something similar for wan2.1. Thanks

SECourses

i plan to make a really good gui to train wan 2.1 but didnt have time yet

Stevie Tee

Hi, I've tested your FramePack Gradio version against a couple of ComfyUI wrapper versions using the same prompts and almost identical settings and your Gradio custom version is the fastest every time. Is this because of how you have optimized it ?

SECourses

i use native code as a base published by the original developer. probably that is why. And we have all the optimizations like Teacache, Flash Attention, Sage Attention, Triton, xFormers all automatically installed into our native Windows Python VENV

Sirdrak

Very happy with the app. I just have one question: Is there a way to use multiple LoRas simultaneously, like in Framepack Studio? Or can only one be used at a time?

SECourses

i am trying to add multi lora for you right now. keep an eye on this post

SteamPull

ERROR DURING GENERATION: CUDA error: an illegal memory access was encountered 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. I am getting this error while running it start the generation but when 1 sec complete this error is showing and generation stop

SECourses

you need to have python 3.10.11 installed in your system. after that delete venv and make new install with latest zip

cool1

Is there any way, with FramePack / FramePack F1 (or something like it) where we could have camera controls to control where the camera moves and how fast? eg. so we can have a camera constantly moving backwards through a street while a person walks forward at the same pace? When I put in the prompt "The camera constantly moves back through the scene at walking pace" it made the camera move forwards and the man walking walk backwards. I've seen camera controls in some online AI image to videos (that usually aren't free) in youtube videos about them. Also I've started gettting the error "ERROR DURING GENERATION: The size of tensor a (67) must match the size of tensor b (66) at non-singleton dimension 3". Is related to image size of duration? That's with target width 960, target height 540, resolution 720. I'm not sure why it needs all 3 (width, height and duration). Isn't width & height enough? The duration was set to 101 seconds. Can't it adjust things automatically so it won't get that error or give more info about it? It says width & height get snapped to the nearest bucket size (so I assume for that it should always set an okay value).

puk

Hy Furkan, when i updating i get this error: Looking in indexes: https://download.pytorch.org/whl/nightly/cu128 ERROR: Could not find a version that satisfies the requirement torch==2.7.0.dev20250311 (from versions: 2.8.0.dev20250318+cu128, 2.8.0.dev20250319+cu128, 2.8.0.dev20250320+cu128, 2.8.0.dev20250321+cu128, 2.8.0.dev20250322+cu128, 2.8.0.dev20250323+cu128, 2.8.0.dev20250324+cu128, 2.8.0.dev20250325+cu128, 2.8.0.dev20250326+cu128, 2.8.0.dev20250327+cu128, 2.8.0.dev20250401+cu128, 2.8.0.dev20250402+cu128, 2.8.0.dev20250403+cu128, 2.8.0.dev20250404+cu128, 2.8.0.dev20250405+cu128, 2.8.0.dev20250406+cu128, 2.8.0.dev20250407+cu128, 2.8.0.dev20250408+cu128, 2.8.0.dev20250409+cu128, 2.8.0.dev20250410+cu128, 2.8.0.dev20250411+cu128, 2.8.0.dev20250412+cu128, 2.8.0.dev20250413+cu128, 2.8.0.dev20250414+cu128, 2.8.0.dev20250415+cu128, 2.8.0.dev20250416+cu128, 2.8.0.dev20250417+cu128, 2.8.0.dev20250418+cu128, 2.8.0.dev20250419+cu128, 2.8.0.dev20250420+cu128, 2.8.0.dev20250421+cu128, 2.8.0.dev20250422+cu128, 2.8.0.dev20250423+cu128, 2.8.0.dev20250424+cu128, 2.8.0.dev20250425+cu128, 2.8.0.dev20250426+cu128, 2.8.0.dev20250427+cu128, 2.8.0.dev20250428+cu128, 2.8.0.dev20250429+cu128, 2.8.0.dev20250430+cu128, 2.8.0.dev20250501+cu128, 2.8.0.dev20250502+cu128, 2.8.0.dev20250503+cu128, 2.8.0.dev20250504+cu128, 2.8.0.dev20250505+cu128, 2.8.0.dev20250506+cu128, 2.8.0.dev20250507+cu128, 2.8.0.dev20250508+cu128, 2.8.0.dev20250509+cu128, 2.8.0.dev20250510+cu128, 2.8.0.dev20250511+cu128, 2.8.0.dev20250512+cu128, 2.8.0.dev20250513+cu128, 2.8.0.dev20250514+cu128, 2.8.0.dev20250515+cu128, 2.8.0.dev20250516+cu128) ERROR: No matching distribution found for torch==2.7.0.dev20250311 The App work after the update. Can i ignore the error massage?

SECourses

you can but it is better to delete venv and run installer again

SECourses

540 cant be divided to 32. that camera etc are depending on the model capabilities :(

SECourses

just added multiple LoRAs. please update test and let me know if works good for you

SECourses

just added multiple LoRAs. please update test and let me know if works good for you

cool1

Thanks. I've changed that width and height to 1024x576 and that's working now (not fast though). With the 540 the field help says "Desired output height. Will be snapped to nearest bucket.". Doesn't that mean it should be snapped to the nearest bucket mean that if you enter a value that isn't divisible by 32 that it should be snapping to a nearest bucket size that is valid (or snapping to the nearest value divisible by 32 when it runs, so that we won't get that (not very clear) error? And if it can't snap to or round to the nearest 32 divisor, couldn't it warn if it's not divisible by 32 instead of continuing and then giving a tensor error?

Sirdrak

A lot of thanks!! It works really really well 😊

koon grizzly

i have been testing camera movements in my last 2 videos, not many seeds work well for camera movement but here are a couple of prompts and seeds i used : zoom in, zoom out, turn inside out, turn around, the camera follows the 'put subject' seeds used : 1111805386 1085772683 1888731491 my newest test videos : https://youtu.be/aQaZ74Fv2E4 https://youtu.be/m8XwJmoSUmk Using timestamps to change camera change works but it is a story of 'test and retest' using same seeds. Like that i was able to change the camera angle in the second video with the alien a couple of times without the need to create a new video part. (by the way, the alien and the space ship was all created with prompts also :-) Easiest is offcourse to create small video parts and combine them yourself or create transitions with start and end frame in original framepack A third one is still generating with only one photo and one prompt where (if everything goes well) it will keep following traffic in a cyberpunk city for a full minute.

koon grizzly

Same here, i have 3 versions of framepack on the pc (thanks to the tip to change location of downloaded models so they all use the same directory) and this version version is way faster then the original framepack and also faster then the framepack studio version. Mostly because i can change latent window size to create bigger video parts and also because we can change teacache settings. For short videos up to 20 seconds the difference in speed is huge, i can have 20 seconds in about 20 minutes with this modded version when creating 3.5 second video parts while it takes up to an hour in original framepack when creating 1 second video parts

koon grizzly

Thanks for another update. When you find the time, can you have a look to give us Temporally metadata files that are created as soon as the video generation starts ? Like this i will always have the seed, even if i interrupt the video or when something goes wrong and the video is not finished. Maybe it can have it's own folder to keep the finished output folder clean ?

cool1

Thanks. The prompt "the camera follows the [eg. person]" seems to work on normal Frame Pack at least it does with the current test I'm doing. It didn't when I tried Frame Pack F1 (F1 seems less likely to properly follow prompts even though that should be the fastest to use if it worked as well as the other). When going down a street (going back from an end frame) Frame Pack also starts getting blurrier for the street background with vertical lines on it (I'm only rendering at 832x480 currently and when I tried with 1024x576 it didn't show the vertical lines but still started getting blurry with some slightly cross hatching). Maybe it's because I'm trying to generate 2 minute videos and shorter videos might work better but it gets blurry in the background walking down the street with the camera following (moving back) even after about 10 secs (so in a 19 sec intermediate vid the first 9 secs have a blurry background with vertical-ish lines at 832x480 res).

koon grizzly

https://youtu.be/2jV9Qp2m3xw Source : Grok ai generated photo of a cyberpunk style city, reduced to 50kb and 854*480 resolution (gives me better results then using original quality) Model: FramePack F1 Prompt: [0] zoom in and follow the traffic Negative Prompt: Seed: 1888731491 TeaCache: Enabled (Threshold: 0.2) Video Length (seconds): 60 FPS: 28 Latent Window Size: 26 Steps: 25 CFG Scale: 1 Distilled CFG Scale: 10 Guidance Rescale: 0 Resolution: 480 Final Width: 640 Final Height: 352 Generation Time: 2 hours 37 min 28 seconds it worked out pretty well i think, to bad you can see the transitions between the parts so easy when creating 3.5 second parts Fun story, it took about 20 minutes for first 20 seconds, an hour to create 42 seconds, and over 90 minutes to get the remaining 20 seconds, it becomes slower with every part, the last 3.5 second video part took almost half an hour :(. (in the beginning it was less then 3 minutes for 3.5 seconds) when it gets to this part : Latents to decode shape: torch.Size([1, 16, 443, 44, 80]) That's when things start to really slow down after a while; my ram is at 50 gigabyte from the 64 at that moment so it can't be that, the gpu show 100% load but it' shows this in the beginning also.

koon grizzly

it doesn't take long for the model to start inventing things, i like that for the things i want to do with it but for serious video work this model needs a lot of improvement. The best you can do to get the best results is to create small video parts, use the last frame (or one of the last frames as it usually slows down in the last couple of frames) and create a new video with a new prompt with that frame, and i also always ad [0] before the prompt so it hopefully starts as soon as the first frame starts

koon grizzly

i will now create a bunch of 20 second 'walking in streets' videos and do some testing with prompts First one looks good : seed 1888731491 prompt [0] walking in the street

NM

hello, after every video generation, why does the consolse always say "press any key.." and shut down for me? I have to restart it after every video

cool1

Thanks. So yours were better than mine for that. So I assume it's generally better to generate closer to 20 second videos and join them/create them starting from the almost last good frame of one of the previous ones than trying to do quite a few time based prompts while trying to generate a 1 minute or 2 minute video. Around 1:43 & 1:56 it seemed to get similar cross-hatching in the tree leaves and some slightly similar vertical lines in the trees around 2:00 and 2:04 (but not as bad as mine were. Maybe the lines and cross-hatching depends on resolution settings and maybe window size. I'm still not sure why in this current UI we have resolution, width & height (when if we have width & height there should be no need for a separate resolution setting). I assume you're using a newer card than a RTX3090 as you seemed to be able to generate that a lot faster than I could.

SECourses

i dont understand your question

koon grizzly

To have a log file as soon as i press generate, so if something goes wrong i can still find the seed and the prompt and settings that were used.

koon grizzly

yes i have most success just creating small 15-30 seconds videos and combining them with last frame to start a new video. (or pick any frame from a video and start over there is even better, but it takes some time to combine them later) I also like using the transition effects in original framepack (+-last frame from a video with a complete new photo and do about 4-8 seconds which often gives a nice transition between the images (as usual many seeds give rubbish results) i'm using an rtx 3080 at the moment, looking for a cheap 3090 or 4080 to speed up things :-) i use 3 or 4 second parts to create the first 20-30 seconds very fast. my settings : TeaCache: Enabled (Threshold: 0.15) i sometimes use 0.1 if my image has much detail on the hands but it takes more time. FPS: 25 Latent Window Size: 26 (= 4 second parts, put at 20 for 3 second video parts) Steps: 25 CFG Scale: 1 Distilled CFG Scale: 10 Guidance Rescale: 0 Resolution: 480 (640 is to high for me, it's easier to upscale later if needed) Final Width: 640 Final Height: 352 (this is set by the software after entering my photo in 854*480 format) You can also put it at 20 steps to go even faster without losing to much movement in easy photos.( it depends a little on the detail in the photo but for simple photo it's about 35 minutes for 30 seconds or 20 minutes for 20 seconds but it goes slower every loop so i try to avoid making more then 7 (28 seconds) or 8 loops (32 seconds)

cool1

Thanks a lot. I've made a note of those settings.

SECourses

ye but that doesnt make sense to generate txt files without video files

cool1

Though we could still end up with intermediate files (but not the final complete video) where we can't see what settings and prompt was used to create those intermediate files. eg. if we stop it or the program ends after it's just done 1 or 2 intermediate files we can't see the prompt and other settings used to create those in a .txt file.

koon grizzly

I almost never finish the video creation, so i never have metadata file and there is no way to find out what seed was used.. :( in framepack studio i get log file as soon as i press generate with all the settings, prompts, seed,...

SECourses

ok with v55 i have added this feature for you. test and let me know

SECourses

ok with v55 i have added this feature for you. test and let me know

koon grizzly

Cool thanks, i will test when i get home from work

koon grizzly

Batch process seems broken for me, gives error and don't want to start, deleted venv directory and new install but problem not fixed. Found 66 images. Sorting naturally... Traceback (most recent call last): File "C:\framepackmodded\FramePack\venv\lib\site-packages\gradio\queueing.py", line 715, in process_events response = await route_utils.call_process_api( File "C:\framepackmodded\FramePack\venv\lib\site-packages\gradio\route_utils.py", line 322, in call_process_api output = await app.get_blocks().process_api( File "C:\framepackmodded\FramePack\venv\lib\site-packages\gradio\blocks.py", line 2137, in process_api result = await self.call_function( File "C:\framepackmodded\FramePack\venv\lib\site-packages\gradio\blocks.py", line 1675, in call_function prediction = await utils.async_iteration(iterator) File "C:\framepackmodded\FramePack\venv\lib\site-packages\gradio\utils.py", line 735, in async_iteration return await anext(iterator) File "C:\framepackmodded\FramePack\venv\lib\site-packages\gradio\utils.py", line 729, in __anext__ return await anyio.to_thread.run_sync( File "C:\framepackmodded\FramePack\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:\framepackmodded\FramePack\venv\lib\site-packages\anyio\_backends\_asyncio.py", line 2470, in run_sync_in_worker_thread return await future File "C:\framepackmodded\FramePack\venv\lib\site-packages\anyio\_backends\_asyncio.py", line 967, in run result = context.run(func, *args) File "C:\framepackmodded\FramePack\venv\lib\site-packages\gradio\utils.py", line 712, in run_sync_iterator_async return next(iterator) File "C:\framepackmodded\FramePack\venv\lib\site-packages\gradio\utils.py", line 873, in gen_wrapper response = next(iterator) File "C:\framepackmodded\FramePack\app.py", line 2373, in batch_process current_adapter_name =currently_loaded_lora_info ["adapter_name"]if currently_loaded_lora_info ["adapter_name"]else "None" TypeError: list indices must be integers or slices, not str

koon grizzly

I hear They fixed the online problem in the new update from framepack studio 0.3, no more need to have internet connection (edit, deleting and reinstalling with your installer also gave the updated version)

TAGopolis

In the latest version, no matter what I put in the prompt, all it does is make my character blink his eyes. That isn't even part of my prompt. Any ideas?

koon grizzly

Have you tested with another photo ? Sometimes a photo don't want to work well for me also

SECourses

yes fixed yesterday sorry for this error - update to latest please

TAGopolis

Will give it a shot. Kinda need it to work with this one but...so frustrated at this point.

TAGopolis

Okay, I used the Leo photo and it....sort of worked. It paid attention to part of the prompt. I am sure I just need to get better at my prompt writing. Appreciate your efforts. Greetings and thanks from NY.

koon grizzly

This is where i originally hear about the update, then i did a google search for framepack studio 0.3 and like that i found the patreon site of that guy but no link to install. So i completely deleted your framepack studio one click install version and reinstalled from scratch and all the updates were inserted after that. https://www.youtube.com/watch?v=ZpKDdh0fT4I

koon grizzly

if i go from text to video, i can add details, even ask grok or chat gpt to create a nice prompt and many things will be followed, but from photo to video i try to keep it as simple as possible and maximum 2 things in 1 prompt. If i add details usually most is skipped and only the last part is somewhat followed

koon grizzly

another tip if it has to be that photo : lower the resolution and/or save in another format (for example from jpg to png)

koon grizzly

Something i don't understand.. In your version, when i raise latent window size, i get longer intermediate parts, (even if i leave FPS at default 30fps). This allows me to create 3 or 4 second video parts instead of 1 second video parts which speeds up creation of a 30 second video from 1 hour to 30-35 minutes (on an rtx3080) In the original Framepack F & framepack studio version, if i raise the latent window size, it will still take an hour for 30 seconds, the intermediate video parts stay at 1 second and i get messed up videos. So you have done something really great with that latent window size that does not work in any other version i tested :-)

TAGopolis

These are great suggestions. I should elaborate. This is a 896x896px png image of me in pixar format. The prompts are: [0] A person plays with gamepad [2] The person gets mad All it generates is me blinking, no matter what I do. I am now doing a fresh install to see if that makes a difference.

SECourses

yes it was a huge effort and improvement :)

TAGopolis

Fresh install yields the same results.

SECourses

looks like base model cant handle your photo probably model limitation

koon grizzly

in v56 the error still came. woke up and it did one video and next this error came : File "C:\framepackmodded\FramePack\venv\lib\site-packages\gradio\queueing.py", line 715, in process_events response = await route_utils.call_process_api( File "C:\framepackmodded\FramePack\venv\lib\site-packages\gradio\route_utils.py", line 322, in call_process_api output = await app.get_blocks().process_api( File "C:\framepackmodded\FramePack\venv\lib\site-packages\gradio\blocks.py", line 2137, in process_api result = await self.call_function( File "C:\framepackmodded\FramePack\venv\lib\site-packages\gradio\blocks.py", line 1675, in call_function prediction = await utils.async_iteration(iterator) File "C:\framepackmodded\FramePack\venv\lib\site-packages\gradio\utils.py", line 735, in async_iteration return await anext(iterator) File "C:\framepackmodded\FramePack\venv\lib\site-packages\gradio\utils.py", line 729, in __anext__ return await anyio.to_thread.run_sync( File "C:\framepackmodded\FramePack\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:\framepackmodded\FramePack\venv\lib\site-packages\anyio\_backends\_asyncio.py", line 2470, in run_sync_in_worker_thread return await future File "C:\framepackmodded\FramePack\venv\lib\site-packages\anyio\_backends\_asyncio.py", line 967, in run result = context.run(func, *args) File "C:\framepackmodded\FramePack\venv\lib\site-packages\gradio\utils.py", line 712, in run_sync_iterator_async return next(iterator) File "C:\framepackmodded\FramePack\venv\lib\site-packages\gradio\utils.py", line 873, in gen_wrapper response = next(iterator) File "C:\framepackmodded\FramePack\app.py", line 2759, in batch_process if current_adapter_name !="None": NameError: name 'current_adapter_name' is not defined ERROR:asyncio:Exception in callback _ProactorBasePipeTransport._call_connection_lost(None) handle: Traceback (most recent call last): File "C:\Python3_10_11\lib\asyncio\events.py", line 80, in _run self._context.run(self._callback, *self._args) File "C:\Python3_10_11\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 Unloaded DynamicSwap_LlamaModel as complete. Unloaded CLIPTextModel as complete. Unloaded SiglipVisionModel as complete. Unloaded AutoencoderKLHunyuanVideo as complete. Unloaded DynamicSwap_HunyuanVideoTransformer3DModelPacked as complete. Worker override: Restored global frame paths.

koon grizzly

looks like it's fixed in v57, running a quick batch test with 3 second videos and it meanwhile started the third video. Thanks, i will try again my list next night :-)

Junior

Hugging Face cache folders set to: D:\FramePack_v53\FramePack\models W0524 16:46:50.058000 30444 venv\Lib\site-packages\torch\distributed\elastic\multiprocessing\redirects.py:29] NOTE: Redirects are currently not supported in Windows or MacOs. Currently enabled native sdp backends: ['flash', 'math', 'mem_efficient', 'cudnn'] Xformers is installed! Flash Attn is installed! Sage Attn is installed! Namespace(share=False) Free VRAM 14.6953125 GB High-VRAM Mode: False Downloading shards: 100%|██████████████████████████████████████████████████████████████| 4/4 [00:00<00:00, 2535.09it/s] Loading checkpoint shards: 100%|█████████████████████████████████████████████████████████| 4/4 [00:00<00:00, 10.49it/s] Loading initial transformer model: Original FramePack Fetching 3 files: 100%|██████████████████████████████████████████████████████████████████████████| 3/3 [00:00 block .launch ( File "D:\FramePack_v53\FramePack\venv\lib\site-packages\gradio\blocks.py", line 2685, in launch raise Exception( Exception: Couldn’t start the app because 'http://127.0.0.1:7860/gradio_api/startup-events' failed (code 502). Check your network or proxy settings to ensure localhost is accessible. 请按任意键继续. . .

koon grizzly

Can somebody confirm they got a working Lora in framepack F ? I have been working a couple of days with framepack studio with a drone shot lora ( https://civitai.com/models/1247109/high-speed-drone-shot-hunyuan-video-lora ) It really works will in there and gives fast motion videos but when i load the lora in this version (same prompt, settings, seed,...) the speed of the camera is the same, as if the lora isn't loaded.

koon grizzly

At the start of the creation it does say the lora has been loaded. so i wonder why it's not doing what it is supposed to do :-) LoRA Change Detected: Target=['high-speed-drone-shot-hunyuan-v1.0-heavy-vfx_ai', None, None, None], Current=['high-speed-drone-shot-hunyuan-v1.0-heavy-vfx_ai', None, None, None] Unloading previous LoRA structures... Unloading LoRA using disable_adapters method Cleared peft_config. Successfully unloaded all previous LoRAs. Loading new LoRA structure: high-speed-drone-shot-hunyuan-v1.0-heavy-vfx_ai from C:\framepackmodded\FramePack\loras\high-speed-drone-shot-hunyuan-v1.0-heavy-vfx_ai.safetensors LoRA weights loaded successfully. Successfully loaded structure for high-speed-drone-shot-hunyuan-v1.0-heavy-vfx_ai. Final currently_loaded_lora_info: [{'adapter_name': 'high-speed-drone-shot-hunyuan-v1.0-heavy-vfx_ai', 'lora_path': 'C:\\framepackmodded\\FramePack\\loras\\high-speed-drone-shot-hunyuan-v1.0-heavy-vfx_ai.safetensors'}, {'adapter_name': None, 'lora_path': None}, {'adapter_name': None, 'lora_path': None}, {'adapter_name': None, 'lora_path': None}]

SECourses

thank you so much fixed with v59

koon grizzly

You got it to work already ? damn, thanks haha !

cool1

edit: I just read your post. Thanks for adding a .txt file the prompt info at the start.

cool1

Can you (or maybe in the app) explain more about how the resolution, target width and target height work? Why does it seem to work differently for FramePack F1 compared to FramePack (standard)? eg. for the last few tries, I've been entering (as far as I know) resolution: 720, target width: 1280, target height 720. Why do we need the extra "resolution" now that we have both the width & height settings? Is it generating it at a lower resolution and then scaling it to the target width & height or something else? When I created a FramePack with those settings (res 720, target width 1280, target height 1280) it created video with that size. The output .txt file says: Model: Original FramePack Resolution: 720 Final Width: 1280 Final Height: 720 so that seems right. The output video is 1280x720 pixels. When I'm created a video in with FramePack F1 I used the same settings I think, but it comes out lower resolution. eg. now I'm creating videos in batch mode with FramePack F1, in the GUI I entered: Resolution 720, Target Width: 1280, Target Height: 720. Each of the image files in the source folder are 1280x720 pixels. but the temporary output metadata file says: Model: FramePack F1 Resolution: 720 Final Width: 960 [why not 1280?] Final Height: 512 [why not 720?] and my intermediate video files are showing as 960x512 (which is 1.875:1). Why isn't it 1280x720 (1.778:1) like my normal FramePack video was? Is FramePack F1 more limited in width & height it can use? How can I get it to output 1280x720 videos when using FramePack F1? I can do it by entering 1280 width & 720 height when doing FramePack videos but it doesn't work for FramePack F1. Maybe it's because one of the values has to be divisible by 32 or some other number (720 isn't evenly divisible by 32) but it doesn't show in the UI which values need to be divisible by what numbers. It would help if it did. But I don't see why 1280x720 worked for FramePack but not for FramePack F1 unless they have different limits on resolutions / different bucket size limits. Ideally there could be a help page in the app that showed what the different limits are. Because the current F1 will generate them at the slightly wrong aspect ratio now even if you specify 1280x720 and it adds black bars when there weren't any in the source image. If I want FramePack F1 videos at 1280x720 should I set the width & height to 1280 × 736 (as both those are divisible by 32), while still using the same 1280x720 source images and then crop any black bars it adds later? edit: I entered 1280x736 (both divisible by 32) in batch mode with FramePack F1 but the output .txt still says Resolution: 720 Final Width: 960 Final Height: 512 Maybe it's because 960 is the max dimension FramePack F1 can do in the buckets. Though now I can't recreate the standard FramePack output where I created an output of 1280x720 and the .txt file said 1280x720. I'm sure it was from the same version of the app because I created it today. In FramePack F1. I assume I need to set "resolution" to a value >= the max resolution of the target width & height (eg. 1320 for 1280x720). It seems to work now but it would help a lot if it explained this in some help in the app. Though if it's going to do all the processing at 1320 width and then downscale it at the end to 1280x720 it might be better to have it generate at 1320 width for higher quality? Also in the output .txt file it says "RIFE FPS Multiplier: 2x FPS" but it doesn't say whether rife was enabled. I hadn't enabled refine for that generation. Also when I'm batch processing where the "use random seed" checkbox was checked, in the output file it said "Use Random Seed Setting: False". Shouldn't that have said "Use Random Seed Setting: True" when that checkbox was checked?

David

Hi! I just got an nvlink bridge for 2x RTX3090. I noticed that while in sli mode, the memory for both video cards fills up equally but the gpu processor usage goes up just for one. And it takes the same time to generate using sli or disabling it (When disabling it, only 1 gpu memory is filled and only that gpu's processor works). Is there a way to take advantage of the full 48gb fo memory using framepack? Haven't tested SwarmUI yet. Thanks!

David

Update: I tried SWARMUI with WAN i2v, it also fills both cards' memories (it looks like it's the same data in both cards since it's exactly the same amount of memory used) but only one works, and then runs out of memory

cool1

Someone says there (in the github link) "As a temporary fix, you can comment this section out if the issue won't be resolved soon.". Will that work if the section of code he mentions is just commented out (wherever it is)? Or might it affect something else (like could it be really needed sometimes?). If it will work can could that be commented out please (I tried running it over night but I forgot that it doesn't work with the router turned off and it gave a connection error soon after. Then it seems to have trouble restarting batch mode after those connection errors. Maybe I'll have to stop and close the app and restart that way.). Or maybe it will be fixed when things are updated to a new version of FramePack (framepack studio 0.3) but if commenting out a small code section could make it work that might be easier. Unless I need to delete all of framepack and re-install and then it will work (but then that will take ages with GB's to download and also all the output files will need moving and restoring).

Tomas Villarreal

Have a way to replace Hanyuan for Wan 2.1? cuz Wan 2.1 give me better results in comfyui, the Hanyuan ones feels like dosnt move too much

SECourses

sadly not because framepack not trained for Wan 2.1 :(

SECourses

SLI can only help you as far as i know if only the APP and model supporting. you will find support mostly from LLMs not from other models sadly

SECourses

disable vpn. probably it is breaking. or bypass somehow

cool1

The latest update says about Loras working - is it going to be possible for us to train a Lora for it or is that unlikely for local windows pcs?

Tomas Villarreal

Have a way to put a new Hunyuan model? like one from civitai?

Salihz

Hello, I got these error with the Windows Install, could you tell me what I did wrong ? Defaulting to user installation because normal site-packages is not writeable ERROR: flash_attn-2.7.4.post1-cp310-cp310-win_amd64.whl is not a supported wheel on this platform. Defaulting to user installation because normal site-packages is not writeable ERROR: deepspeed-0.16.5-cp310-cp310-win_amd64.whl is not a supported wheel on this platform. Defaulting to user installation because normal site-packages is not writeable ERROR: sageattention-2.1.1+cu128torch2.7.0-cp310-cp310-win_amd64.whl is not a supported wheel on this platform. Defaulting to user installation because normal site-packages is not writeable ERROR: xformers-0.0.30+3abeaa9e.d20250424-cp310-cp310-win_amd64.whl is not a supported wheel on this platform.

SECourses

yes your error is you dont have python 3.10.11 installed. you have to have it.

SECourses

hi files are saved inside tmp inside installed folder. you can delete it anytime. 16 GB works perfect. just make sure to have virtual RAM set. like 64 GB virtual ram

SECourses

i didnt put into downloader yet but loras working great.

Salihz

Thanks a lot ! thought the lastest was enough

SECourses

with python and AI latest is never best :D i learnt this hard truth in years

koon grizzly

Is anyone with a RTX 3090 able to undervolt AND use framepack? Even the default undervolt settings (850mv - 1800mhz) make the card crash as soon as i use framepack, i even tested 1750mhz I ran long tests, played games for a couple of hours that took the card all the time to 99% load. Everything runs perfect. Until i start framepack. When the card is at default everything works fine but that card goes up to 390 watts and my pc used 21.5kwu this weekend lol This is the error i get in Framepack : ERROR DURING GENERATION: CUDA error: an illegal memory access was encountered 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. ================================================== Traceback (most recent call last): File "C:\framepackmodded\FramePack\app.py", line 1575, in worker generated_latents =sample_hunyuan ( File "C:\framepackmodded\FramePack\venv\lib\site-packages\torch\utils\_contextlib.py", line 116, in decorate_context return func(*args, **kwargs) File "C:\framepackmodded\FramePack\diffusers_helper\pipelines\k_diffusion_hunyuan.py", line 116, in sample_hunyuan results = sample_unipc(k_model, latents, sigmas, extra_args=sampler_kwargs, disable=False, callback=callback) File "C:\framepackmodded\FramePack\diffusers_helper\k_diffusion\uni_pc_fm.py", line 141, in sample_unipc return FlowMatchUniPC(model, extra_args=extra_args, variant=variant).sample(noise, sigmas=sigmas, callback=callback, disable_pbar=disable) File "C:\framepackmodded\FramePack\diffusers_helper\k_diffusion\uni_pc_fm.py", line 118, in sample model_prev_list = [self.model_fn(x, vec_t)] File "C:\framepackmodded\FramePack\diffusers_helper\k_diffusion\uni_pc_fm.py", line 23, in model_fn return self.model(x, t, **self.extra_args) File "C:\framepackmodded\FramePack\diffusers_helper\k_diffusion\wrapper.py", line 37, in k_model pred_positive = transformer(hidden_states=hidden_states, timestep=timestep, return_dict=False, **extra_args['positive'])[0].float() File "C:\framepackmodded\FramePack\venv\lib\site-packages\torch\nn\modules\module.py", line 1751, in _wrapped_call_impl return self._call_impl(*args, **kwargs) File "C:\framepackmodded\FramePack\venv\lib\site-packages\torch\nn\modules\module.py", line 1762, in _call_impl return forward_call(*args, **kwargs) File "C:\framepackmodded\FramePack\diffusers_helper\models\hunyuan_video_packed.py", line 986, in forward hidden_states, encoder_hidden_states = self.gradient_checkpointing_method( File "C:\framepackmodded\FramePack\diffusers_helper\models\hunyuan_video_packed.py", line 836, in gradient_checkpointing_method result = block(*args) File "C:\framepackmodded\FramePack\venv\lib\site-packages\torch\nn\modules\module.py", line 1751, in _wrapped_call_impl return self._call_impl(*args, **kwargs) File "C:\framepackmodded\FramePack\venv\lib\site-packages\torch\nn\modules\module.py", line 1762, in _call_impl return forward_call(*args, **kwargs) File "C:\framepackmodded\FramePack\diffusers_helper\models\hunyuan_video_packed.py", line 598, in forward hidden_states = gate * self.proj_out(hidden_states) File "C:\framepackmodded\FramePack\venv\lib\site-packages\torch\nn\modules\module.py", line 1751, in _wrapped_call_impl return self._call_impl(*args, **kwargs) File "C:\framepackmodded\FramePack\venv\lib\site-packages\torch\nn\modules\module.py", line 1762, in _call_impl return forward_call(*args, **kwargs) File "C:\framepackmodded\FramePack\venv\lib\site-packages\peft\tuners\lora\layer.py", line 712, in forward result = self.base_layer(x, *args, **kwargs) File "C:\framepackmodded\FramePack\venv\lib\site-packages\torch\nn\modules\module.py", line 1751, in _wrapped_call_impl return self._call_impl(*args, **kwargs) File "C:\framepackmodded\FramePack\venv\lib\site-packages\torch\nn\modules\module.py", line 1762, in _call_impl return forward_call(*args, **kwargs) File "C:\framepackmodded\FramePack\venv\lib\site-packages\torch\nn\modules\linear.py", line 125, in forward return F.linear(input, self.weight, self.bias) File "C:\framepackmodded\FramePack\diffusers_helper\memory.py", line 26, in hacked_get_attr return torch.nn.Parameter(p.to(**kwargs), requires_grad=p.requires_grad) RuntimeError: CUDA error: an illegal memory access was encountered 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. ================================================== Error during memory cleanup after exception: CUDA error: an illegal memory access was encountered 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.

koon grizzly

could it be this easy ? i left everything default in MSI afterburner but lowered the maximum power it can have from 100% to 93% (did some testing last couple of days) Now the 3090 runs between 270 and 312 (peak) watts in framepack. Yes i lost a little bit of power, it takes a couple of extra seconds to get a second of footage, but a couple of minutes extra for 120 seconds of video don't really matter. For gaming, in the benchmark included in Cyberpunk i only lost 0.9 average fps (highest settings) so not noticeable. Temps are also down a lot, now runs between 55 and 63 (peak) Celsius while it went up to 67 before the testing

SECourses

yep that easy : "an illegal memory access was encountered"

koon grizzly

No more errors now, i can even put it at 80% when i go to bed and videos are created when i wake up without errors

Lawren

Error from FramePack_v59 most likely due to silent updates to dependencies: Ubuntu@0328-kci2-prxmx100090:~/Downloads/FramePack_v59$ export PYTHONWARNINGS=ignore cd FramePack source ./venv/bin/activate export HF_HUB_ENABLE_HF_TRANSFER=1 python3 app.py --share Traceback (most recent call last): File "/home/Ubuntu/Downloads/FramePack_v59/FramePack/venv/lib/python3.10/site-packages/transformers/utils/import_utils.py", line 1778, in _get_module return importlib.import_module("." + module_name, self.__name__) File "/home/Ubuntu/.pyenv/versions/3.10.16/lib/python3.10/importlib/__init__.py", line 126, in import_module return _bootstrap._gcd_import(name[level:], package, level) File "", line 1050, in _gcd_import File "", line 1027, in _find_and_load File "", line 1006, in _find_and_load_unlocked File "", line 688, in _load_unlocked File "", line 883, in exec_module File "", line 241, in _call_with_frames_removed File "/home/Ubuntu/Downloads/FramePack_v59/FramePack/venv/lib/python3.10/site-packages/transformers/models/llama/modeling_llama.py", line 32, in from ...modeling_flash_attention_utils import _flash_attention_forward File "/home/Ubuntu/Downloads/FramePack_v59/FramePack/venv/lib/python3.10/site-packages/transformers/modeling_flash_attention_utils.py", line 27, in from flash_attn.bert_padding import index_first_axis, pad_input, unpad_input # noqa File "/home/Ubuntu/Downloads/FramePack_v59/FramePack/venv/lib/python3.10/site-packages/flash_attn/__init__.py", line 3, in from flash_attn.flash_attn_interface import ( File "/home/Ubuntu/Downloads/FramePack_v59/FramePack/venv/lib/python3.10/site-packages/flash_attn/flash_attn_interface.py", line 15, in import flash_attn_2_cuda as flash_attn_gpu ImportError: /home/Ubuntu/Downloads/FramePack_v59/FramePack/venv/lib/python3.10/site-packages/flash_attn_2_cuda.cpython-310-x86_64-linux-gnu.so: undefined symbol: _ZN3c104cuda9SetDeviceEa The above exception was the direct cause of the following exception: Traceback (most recent call last): File "/home/Ubuntu/Downloads/FramePack_v59/FramePack/app.py", line 30, in from transformers import LlamaModel ,CLIPTextModel ,LlamaTokenizerFast ,CLIPTokenizer File "", line 1075, in _handle_fromlist File "/home/Ubuntu/Downloads/FramePack_v59/FramePack/venv/lib/python3.10/site-packages/transformers/utils/import_utils.py", line 1767, in __getattr__ value = getattr(module, name) File "/home/Ubuntu/Downloads/FramePack_v59/FramePack/venv/lib/python3.10/site-packages/transformers/utils/import_utils.py", line 1766, in __getattr__ module = self._get_module(self._class_to_module[name]) File "/home/Ubuntu/Downloads/FramePack_v59/FramePack/venv/lib/python3.10/site-packages/transformers/utils/import_utils.py", line 1780, in _get_module raise RuntimeError( RuntimeError: Failed to import transformers.models.llama.modeling_llama because of the following error (look up to see its traceback): /home/Ubuntu/Downloads/FramePack_v59/FramePack/venv/lib/python3.10/site-packages/flash_attn_2_cuda.cpython-310-x86_64-linux-gnu.so: undefined symbol: _ZN3c104cuda9SetDeviceEa (venv) Ubuntu@0328-kci2-prxmx100090:~/Downloads/FramePack_v59/FramePack$

Lawren

This is VM Massed Compute Install Shell script.

Lawren

Claude's explanation to "temporarily" fix: # FramePack v59 Installation Issue - August 2025 ## Problem Summary The FramePack v59 installation script that worked for months is now failing with import errors, specifically: - `undefined symbol: _ZN3c104cuda9SetDeviceEa` (flash-attn error) - `No module named 'torch._inductor.kernel.flex_attention'` (torchao error) ## Root Cause The script installs **PyTorch 2.7.0.dev20250311** (March 11, 2025 nightly build), but PyTorch automatically removes old nightly builds after 2-3 months. Since it's now August 2025, that specific build is **no longer available**. When pip can't find the requested version, it installs the **current nightly build** instead, which has different internal APIs. The pre-compiled flash-attn and other packages were built against the March version and are incompatible with the current nightly. ## Why This Happened Now - March 11, 2025 nightly build was removed from PyTorch servers (standard cleanup) - Script now gets a different PyTorch version than intended - Pre-compiled wheels (flash-attn, sageattention) no longer match PyTorch version - Dependency cascade failure: torchao → accelerate → diffusers → app crash ## Current Workaround 1. **Remove torchao** (causes most compatibility issues): ```bash pip uninstall torchao -y ``` 2. **Use standard flash-attn** (loses custom optimizations): ```bash pip uninstall flash-attn sageattention -y pip install flash-attn --no-cache-dir ``` ## Long-term Solution Needed The installation script needs updating to use either: - **All stable versions** (PyTorch 2.4.x + compatible packages) - **All current nightlies** (compile everything from source) - **Pinned stable environment** (avoid nightly builds entirely) The current approach of mixing March 2025 nightly ecosystem with August 2025 builds will continue to break as time passes. ## Impact - App functionality works but with reduced performance - No optimized attention mechanisms - Potential compatibility issues with newer models

SECourses

thanks checking right now to fix - massed compute secourses image

Lawren

Claude almost had the explanation right. First I had to run this: deactivate cd ~/Downloads/FramePack_v59 rm -rf FramePack/venv cd FramePack python3 -m venv venv source ./venv/bin/activate Then continue with: pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu128 pip install flash-attn --no-cache-dir pip install deepspeed triton wheel torchao ninja pip install -r requirements.txt Next: pip uninstall torchao -y pip install hf_transfer Then run your start line: export PYTHONWARNINGS=ignore cd FramePack source ./venv/bin/activate export HF_HUB_ENABLE_HF_TRANSFER=1 python3 app.py --share I'm using it without attn mechanisms right now. With DGX A100 you can barely tell the difference.

SECourses

thanks for info. sorry i couldnt test last day. today testing and fixing hopefully

SECourses

hello again. fixed your error with v60. sorry for this. the torch version urls were gone and now i improved process significantly. please read v60 logs

SECourses

hello again. fixed your error with v60. sorry for this. the torch version urls were gone and now i improved process significantly. please read v60 logs

Lawren

No worries, I just wanted to give you a heads up with a small informational jump start. I really appreciate all the work you do for your subscribers by simplifying AI development. Thank you, sir.

Lawren

I can't wait for FramePack P1. These tools are advancing so quickly! Do you plan to revamp your installs once P1 is released?

Lawren

FramePack P1 is still currently in development by lllyasviel. They've got a page that includes a rundown of anti-drift and prediction first results. F1 introduced a lot error accumulation, this new version aims to minimize it. https://lllyasviel.github.io/frame_pack_gitpage/p1/

SECourses

great ty. hopefully i will support it as soon as published

Lawren

That would be wonderful. Thanks, Furkan.

Lawren

Hey Furkan, Quick question for you.. If I rent two GPUs from Massed Compute, can you assign both of them to a single instance of FramePack?

SECourses

nope sadly. you need to run 2 instance of FramePack

Lawren

That's what I thought. Okay, thank you.

Hockey

Is it possible to use this with Wan 2.2?

Bjoern

Thanks for the script. I have a general question: could you make your package on MassedComput a bit leaner? Most users will probably just run the scripts from Patreon and don’t need all the preinstalled programs. Because there’s so much preinstalled, setting up a new virtual environment often takes around 15 minutes. When you’re renting a large GPU, just initializing the machine can already cost nearly $1 — before it’s even fully booted.

SECourses

yes we are working on that trying to reduce its size at the moment

Julius

Thanks for this update.

Julius

Disappointed with the update, the time to generate a video has doubled compared to version 60! Any idea why this has increased?

SECourses

i just compared older installation vs newer installation. 34 vs 36 seconds for first second of same run. no speed difference

Julius

For my test, the same image, the same settings 10 minutes per second against 5 previously (Target Width 1440-676). Did I make a mistake when applying the update?

SECourses

which gpu i dont think so. can you monitor vram and ram of both install?

Julius

Ok, I ran Windows Install and Windows Update again, and it seems like things are back to normal.

SECourses

hi is this comfyui? please give morre info

Christoph Behrmann

I am sorry, it is ComfyUI not SwarmUI, I mixed posts, my fault

Christoph Behrmann

wow, I just found out that you added multiple output resolutions for video gen with up to 1080... I have not tested yet, but are you sure that lllyasviel/FramePackI2V_HY works for more than 480p just because it is build on 480p base model? (I am looking since 2 weeks for some solution to make FramePack or FramePack Studio work with larger video output than 480p, that is why I was going more deep into lllyasviel/FramePackI2V_HY AI model)

SECourses

yes it works up to a degree give it a try

Christoph Behrmann

running with 720 at the moment with my 3090 -- and it is talking its time... usually I have 480x832 and then upscaling (inside FramePack Studio) -- what would you recommend? better to upscale later than generate directly as for example FHD (1080x1920) ? 540x960 would be enough to upscale x2 later for FHD... but 540 is not there

SECourses

480x832 is its original resolution. you can try 1 upper resolution

jacky0409

Since yesterday, Framepack won't launch and I'm getting an error. Hugging Face cache folders set to: E:\FramePack_v61\FramePack\models Skipping import of cpp extensions due to incompatible torch version 2.8.0+cu129 for torchao version 0.14.1 Please see https://github.com/pytorch/ao/issues/2919 for more info W1029 18:25:00.089000 13200 venv\Lib\site-packages\torch\distributed\elastic\multiprocessing\redirects.py:29] NOTE: Redirects are currently not supported in Windows or MacOs. Currently enabled native sdp backends: ['flash', 'math', 'mem_efficient', 'cudnn'] Xformers is installed! Flash Attn is installed! Sage Attn is installed! Namespace(share=False) Free VRAM 10.7919921875 GB High-VRAM Mode: False Loading checkpoint shards: 100%|█████████████████████████████████████████████████████████| 4/4 [00:00<00:00, 8.53it/s] Loading initial transformer model: Original FramePack Loading checkpoint shards: 33%|███████████████████ | 1/3 [00:00<00:00, 8.01it/s] It stops at the execution stage and says, "Press any button to continue." Why is this happening? How can I fix it? I'm using v61, a completely fresh install.

SECourses

hello this is out of RAM. if your virtual disk is full it might be why. your disk space is full? did you set 100 GB virtual ram?

jacky0409

Virtual memory is set to automatic and is currently 15276 MB.

jacky0409

The memory RAM capacity is 32 GB.

SECourses

yep set it to 100 gb restart pc and try and let me know

jacky0409

Hugging Face cache folders set to: E:\FramePack_v61\FramePack\models Skipping import of cpp extensions due to incompatible torch version 2.8.0+cu129 for torchao version 0.14.1 Please see https://github.com/pytorch/ao/issues/2919 for more info W1029 18:59:41.680000 15040 venv\Lib\site-packages\torch\distributed\elastic\multiprocessing\redirects.py:29] NOTE: Redirects are currently not supported in Windows or MacOs. Currently enabled native sdp backends: ['flash', 'math', 'mem_efficient', 'cudnn'] Xformers is installed! Flash Attn is installed! Sage Attn is installed! Namespace(share=False) Free VRAM 10.7919921875 GB High-VRAM Mode: False Loading checkpoint shards: 100%|█████████████████████████████████████████████████████████| 4/4 [00:00<00:00, 8.26it/s] Loading initial transformer model: Original FramePack Loading checkpoint shards: 33%|███████████████████ | 1/3 [00:00<00:00, 8.01it/s]

jacky0409

I'm getting an error even though I set the virtual memory to 100 GB.

jacky0409

It's resolved. I reduced the space on the C: drive and it now runs normally. Thank you for your help.

SECourses

yep exactly as i have guessed :D

Lawren

Peft now upgrades to 0.18.0 which breaks the install. One line change in requirements fixed it. From peft To peft==0.17.0

SECourses

thank you so much for reporting. updated the zip file with this requirements change

Lawren

Hey, Furkan. Do you have any idea what's going on with Massed Compute? Their RTX 6000 Blackwell has been offline for a week with zero response from email contact.

SECourses

just forwarded your message to them

Lawren

Thank you, I really hope it's back online again soon.

SECourses

yes they replied from our discord "Don't recall any emails about blackwells recently. but we are planning for more to be available before the year ends "

Lawren

I sent a generic contact form message from their website, and also emailed Nic Baughman @ Massed, but I'm glad to see their working to provide more availability. Thanks again, Furkan Lawren

Christoph Behrmann

any chance to add hunyuan 1.5 for framepack from your end?

SECourses

what you mean can you eloborate?

Taiga

tried installing this (v62) three times on a windows setup that is set up specifically for all of your other installs: Edit this file and set set HF_HUB_ENABLE_HF_TRANSFER=0 if you get model download errors of HF_TRANSFER Hugging Face cache folders set to: D:\FramePack\FramePack\models Skipping import of cpp extensions due to incompatible torch version 2.8.0+cu129 for torchao version 0.16.0 Please see https://github.com/pytorch/ao/issues/2919 for more info W0307 19:01:20.004000 311588 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 "D:\FramePack\FramePack\venv\lib\site-packages\diffusers\utils\import_utils.py", line 1010, in _get_module return importlib.import_module("." + module_name, self.__name__) File "C:\Python310\lib\importlib\__init__.py", line 126, in import_module return _bootstrap._gcd_import(name[level:], package, level) File "", line 1050, in _gcd_import File "", line 1027, in _find_and_load File "", line 992, in _find_and_load_unlocked File "", line 241, in _call_with_frames_removed File "", line 1050, in _gcd_import File "", line 1027, in _find_and_load File "", line 1006, in _find_and_load_unlocked File "", line 688, in _load_unlocked File "", line 883, in exec_module File "", line 241, in _call_with_frames_removed File "D:\FramePack\FramePack\venv\lib\site-packages\diffusers\models\autoencoders\__init__.py", line 21, in from .autoencoder_rae import AutoencoderRAE File "D:\FramePack\FramePack\venv\lib\site-packages\diffusers\models\autoencoders\autoencoder_rae.py", line 31, in from transformers import ( ImportError: cannot import name 'Dinov2WithRegistersConfig' from 'transformers' (D:\FramePack\FramePack\venv\lib\site-packages\transformers\__init__.py) The above exception was the direct cause of the following exception: Traceback (most recent call last): File "D:\FramePack\FramePack\app.py", line 69, in from diffusers import AutoencoderKLHunyuanVideo File "", line 1075, in _handle_fromlist File "D:\FramePack\FramePack\venv\lib\site-packages\diffusers\utils\import_utils.py", line 1001, in __getattr__ value = getattr(module, name) File "D:\FramePack\FramePack\venv\lib\site-packages\diffusers\utils\import_utils.py", line 1000, in __getattr__ module = self._get_module(self._class_to_module[name]) File "D:\FramePack\FramePack\venv\lib\site-packages\diffusers\utils\import_utils.py", line 1012, in _get_module raise RuntimeError( RuntimeError: Failed to import diffusers.models.autoencoders.autoencoder_kl_hunyuan_video because of the following error (look up to see its traceback): cannot import name 'Dinov2WithRegistersConfig' from 'transformers' (D:\FramePack\FramePack\venv\lib\site-packages\transformers\__init__.py) Press any key to continue . . .

達 _

Hugging Face cache folders set to: E:\python\FramePack_v62\FramePack\models Skipping import of cpp extensions due to incompatible torch version 2.8.0+cu129 for torchao version 0.16.0 Please see https://github.com/pytorch/ao/issues/2919 for more info W0307 22:52:22.277000 3332 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 "E:\python\FramePack_v62\FramePack\venv\lib\site-packages\diffusers\utils\import_utils.py", line 1010, in _get_module return importlib.import_module("." + module_name, self.__name__) File "C:\Program Files\Python310\lib\importlib\__init__.py", line 126, in import_module return _bootstrap._gcd_import(name[level:], package, level) File "", line 1050, in _gcd_import File "", line 1027, in _find_and_load File "", line 992, in _find_and_load_unlocked File "", line 241, in _call_with_frames_removed File "", line 1050, in _gcd_import File "", line 1027, in _find_and_load File "", line 1006, in _find_and_load_unlocked File "", line 688, in _load_unlocked File "", line 883, in exec_module File "", line 241, in _call_with_frames_removed File "E:\python\FramePack_v62\FramePack\venv\lib\site-packages\diffusers\models\autoencoders\__init__.py", line 21, in from .autoencoder_rae import AutoencoderRAE File "E:\python\FramePack_v62\FramePack\venv\lib\site-packages\diffusers\models\autoencoders\autoencoder_rae.py", line 31, in from transformers import ( ImportError: cannot import name 'Dinov2WithRegistersConfig' from 'transformers' (E:\python\FramePack_v62\FramePack\venv\lib\site-packages\transformers\__init__.py) The above exception was the direct cause of the following exception: Traceback (most recent call last): File "E:\python\FramePack_v62\FramePack\app.py", line 69, in from diffusers import AutoencoderKLHunyuanVideo File "", line 1075, in _handle_fromlist File "E:\python\FramePack_v62\FramePack\venv\lib\site-packages\diffusers\utils\import_utils.py", line 1001, in __getattr__ value = getattr(module, name) File "E:\python\FramePack_v62\FramePack\venv\lib\site-packages\diffusers\utils\import_utils.py", line 1000, in __getattr__ module = self._get_module(self._class_to_module[name]) File "E:\python\FramePack_v62\FramePack\venv\lib\site-packages\diffusers\utils\import_utils.py", line 1012, in _get_module raise RuntimeError( RuntimeError: Failed to import diffusers.models.autoencoders.autoencoder_kl_hunyuan_video because of the following error (look up to see its traceback): cannot import name 'Dinov2WithRegistersConfig' from 'transformers' (E:\python\FramePack_v62\FramePack\venv\lib\site-packages\transformers\__init__.py) 請按任意鍵繼續 . . .

SECourses

thanks for reporting. get latest v63 zip file. overwrite all. delete venv and run installer again. i fixed errors and upgraded installers to 100x faster uv

SECourses

thanks for reporting. get latest v63 zip file. overwrite all. delete venv and run installer again. i fixed errors and upgraded installers to 100x faster uv

jacky0409

Are you no longer updating the framepack? It's a shame that it's not being updated anymore.

SECourses

i may update what feature you looking for? not working? there werent any newer framepack model sadly

jacky0409

I wasn't really looking for a specific feature; I was just expressing my disappointment that it isn't being updated anymore. It’s a shame that Framepack feels like an AI generation tool that is practically disappearing now. In my opinion, Framepack is the most accessible video production tool. Still, just knowing that you said you will continue to update it makes me feel a bit less disappointed. Since I am Korean and am using Google Translate to send this message, I am not sure if the meaning is being conveyed clearly. Also, if there is one thing I wish for, it would be that the prompt could be available in Korean, like Grok. That might be difficult given the circumstances, though. I just wanted to mention a wish. Anyway, thank you for answering my question. I hope you stay healthy and have many good things happen to you. Thank you.

SECourses

Thanks. By the way LTX 2.3 also great and I am about to update swarmui post. I recommend enable notifications and check when published. 5 second video takes 24 second to generate with new HQ Int8 ConvRot model