Home
Creators
Search Recent Random
Posts
Search DMs Popular Hash Lookup Tags Random
Importer
Import FAQ
Register Favorites Keys Review DMs Logout
Community
Telegram ThePornDude
Home Creators Posts Import Register Favorites Logout

1-Click INSTALL VACE VIDEO CONTROLNET IMAGE TO VIDEO ComfyUI WebUI! (Patreon)

Published:
2025-05-22 17:24:59
Edited:
2025-05-22 17:25:48
Imported:
2026-06
Tags:

Downloads

Content

Hey everyone! I've created a 1-click installer for the newest "THE NEW AI VIDEO CONTROLNET KING! ", called "VACE". Simply THE MOST POWERFUL UPGRADE TO AI VIDEO GENERATION EVER! A ControlNet-style adapter for WAN 2.1 that gives you perfect control over your video generations using OpenPose, Depth, Segmentation AND WHEN Paired with the special LoRA CausVid, you can now generate videos up to 4x faster than before… It's just amazing!You can check out the video right here: https://youtu.be/zhNKpdL03PAThe workflow: VACE VIDEO CONTROLNET IMAGE TO VIDEO WORKFLOW! | PatreonThe installer of course automates the entire install process for comfyUI and the models downloads!

LOCAL INSTALL!!! RECOMMENDED TO START FROM SCRATCH SINCE YOU NEED THE LATEST COMFYUI TORCH VERSION!!!1. Download the VACE-COMFYUI-MANAGER_AUTO_INSTALL.bat

2. Run the bat file and select either option 1, 2 or 3 depending on your GPU vram

3. Wait for the install to finish

4. The webui will launch, load the VACE_IMAGE-TO-VIDEO_CONTROLNET.json file and there you go😎

IF YOU ALREADY HAVE COMFYUI INSTALLED USING A PREVIOUS INSTALLER:

1. Download the VACE-MODELS-NODES_INSTALL.bat and place it in the ComfyUI_windows_portable\ComfyUI folder

2. Run the bat file and select either option 1, 2 or 3 depending on your GPU vram

3. Wait for the install to finish

4. ???

5. Profit 😎IF YOU ARE USING RUNPOD:

1. Create an account if you haven't already: Runpod

2. Click on Pod (on the left side) then click deploy

3. Choose a GPU with at least 24gb of VRAM (a cheap 3090, 4090 or an A30 are great), click here for the comfyui template (aitrepreneur/comfyui), then edit the template and choose 80gb for both the container and volume disk, then deploy on demand

4. Go to my pods, wait for everything to finish and then click "connect", then "Connect to HTTP SERVICE port 8888"

5. Then go inside the comfyui folder and then drag and drop the HIDREAM_AUTO_INSTALL-RUNPOD.sh file on the left side of the UI then click on the "Terminal" icon on the right side on the UI

6. Copy and paste these two lines then press enter:chmod +x VACE_AUTO_INSTALL-RUNPOD.sh./VACE_AUTO_INSTALL-RUNPOD.sh7. Wait for everything to be installed

8. Go back to the port 3000 and click on the manager to update comfyui and restart it

9. Load the VACE_IMAGE-TO-VIDEO_CONTROLNET.json file and there you go😎Also use the command: tail -f comfyui.log

inside the workspace/logs to get the real time logs so that you always know when the install is done, It will then ask you to restart, so click on that restart button, wait for the webui to reboot, click refresh once it's done and there you go.IF YOU ARE USING RUNPOD AND YOU WANT TO STOP THE POD THEN RESTART IT:1. Go inside your pod2. SERVICE port 88883. Go inside the Comfyui folder, open a terminal icon and re copy and paste those two lines:chmod +x VACE_AUTO_INSTALL-RUNPOD.sh./VACE_AUTO_INSTALL-RUNPOD.sh4. And then in a few seconds it will reinstall all the requirements for the nodes5. Go back to comfyui port 3000, click on manager, click update all, wait for everything to be updated then restart.6. Re drag and drop the VACE_IMAGE-TO-VIDEO_CONTROLNET.json file inside and there you go😎

As always, supporting me on Patreon allows me to keep creating helpful resources like this for the community. Thank you for your support - now go have some fun😉!

Files

Comments

Abdelhakim IRBAH

Hello, thank you for this new flow. I have a question though. When I execute the flow, I only get a single image, not a video. I found this in the logs : Failed to validate prompt for output 8: * LoaderGGUF 21: - Value not in list: gguf_name: 'Wan2.1-VACE-14B-Q4_K_S.gguf' not in ['Wan2.1-VACE-14B-Q8_0.gguf'] Output will be ignored xFormers not available I use a RTX 6000 FYI, so I clicked on the third choice when I installed the package. Do I need to download something missing?

Bobépine

I don't know why I keep on paying for those crappy bat files that are not working. I'm an idiot!!!

Abdelhakim IRBAH

actually it works. the name in the GGUF Loader was not the same as the file. It's OK now. I have another question. I find it very long to generate one video. I don't know enough about dev to understand the logs, but is it possible that it takes too long because my CPU is used instead of my GPU? He are my logs if someone has an idea. Using scaled fp8: fp8 matrix mult: False, scale input: False Requested to load WanTEModel loaded completely 9.5367431640625e+25 6419.477203369141 True CLIP/text encoder model load device: cuda:0, offload device: cpu, current: cuda:0, dtype: torch.float16 C:\Users\Shadow\Downloads\ComfyUI_windows_portable\ComfyUI\custom_nodes\gguf\pig.py:342: UserWarning: The given NumPy array is not writable, and PyTorch does not support non-writable tensors. This means writing to this tensor will result in undefined behavior. You may want to copy the array to protect its data or make it writable before converting it to a tensor. This type of warning will be suppressed for the rest of this program. (Triggered internally at C:\actions-runner\_work\pytorch\pytorch\pytorch\torch\csrc\utils\tensor_numpy.cpp:209.) torch_tensor = torch.from_numpy(tensor.data) gguf qtypes: F32 (836), Q8_0 (489), F16 (6) model weight dtype torch.float16, manual cast: None model_type FLOW Requested to load WanTEModel loaded completely 20036.577736663818 6419.477203369141 True Requested to load WanVAE loaded completely 10479.22240447998 242.02829551696777 True Requested to load WAN21_Vace loaded partially 13552.705709960937 13552.697265625 0 Attempting to release mmap (381)

John__Camp

Yeah, I find there is always something missing and no instructions on how to actually get EVERYTHING working. I can't in good conscience spend another penny on these workflows that don't work and don't flow. I'm sure it has something to do with the fact that I don't really know how to use comfy ui all that well. It's appealing to a novice to see "one click install" like its going to be easy and user-friendly. Without a full list of steps or knowledge to properly troubleshoot, it's a waste for me. There are decent options out there if you're not aiming to make NSFW content.

MegaChill

it's the typical kind of - it works on my machine kinda thing but then selling it

Brian Monarch

2 questions :) 1. I had it working. Made a few videos. I didn't change anything, and all of a sudden the DWPreprocessor is outlined in red (it says the missing node is "comfyui_controlnet_aux" in manager) and also in manager it says "5 node conflicts." So not sure if that's the issue or what but I can't get the red outline of the DWPreprocessor to go away so i can use it now. It worked at first. I tried changing the version to "latest" and the higher numbered ones and the node still won't come back. 2. Can I use the full 14B model? I have a Blackwell. The GGUF node doesn't seem to allow that to be pulled up since it's not gguf I guess. Could I replace the node with a different model loader? Thanks!!!

Morpheus

All nodes installed perfectly but i get this error when trying to run : [Errno 2] No such file or directory: 'C:\\Users\\DDD\\Downloads\\ComfyUI_windows_portable\\ComfyUI\\custom_nodes\\comfyui_controlnet_aux\\ckpts\\depth-anything\\Depth-Anything-V2-Large\\.cache\\huggingface\\download\\EheDQYkKfRTio82VDltlF5Af82M=.a7ea19fa0ed99244e67b624c72b8580b7e9553043245905be58796a608eb9345.incomplete'

Gabbis85

Which VACE quantization? 1) Q4_K_S ÔÇô GPUs < 12 GB vRAM 2) Q5_K_S ÔÇô GPUs 12-16 GB 3) Q8_0 ÔÇô Best quality, GPUs 16 GB+ Enter 1, 2 or 3: 2 -------- Checking prerequisites -------- -------- Downloading ComfyUI -------- % Total % Received % Xferd Average Speed Time Time Time Current Dload Upload Total Spent Left Speed 0 0 0 0 0 0 0 0 --:--:-- --:--:-- --:--:-- 0 100 1992M 100 1992M 0 0 87.9M 0 0:00:22 0:00:22 --:--:-- 105M -------- Extracting ComfyUI -------- Extraction failed. Press any key to continue . . . it did ask before to install 7-zip. so i did download from the orginal 7-zip page and installed it, then i could continue but i get this error above, do i have to install z-zip in another way, help plz ----> yoooo i manage to fix it, just use a simple name on the folder where the install files are located, dont use "!" on the folder :)

Proteinique

yo i didn't understnd how u fixed that, i have the same problem

Gabbis85

hi it first did ask me to install 7-zip to extrakt so i did download it from the 7-zip homepage, https://www.7-zip.org then i installed it, but then i did get the error above "Extraction failed" so i just deleted the corrupt folder where some files where installed to, then i started over again, i just changed the name on the folder where i did have install files located these: VACE_AUTO_INSTALL-RUNPOD.sh VACE-COMFYUI-MANAGER_AUTO_INSTALL.bat VACE-MODELS-NODES_INSTALL.bat ,dont use "!" on the folder. worked without it. hope it helps :)

TinyDevilRay

Noob questions here.... I have 12GB VRAM on my local (slow gpu) and my generations take forever. If I use runpod and select a GPU with 24BG VRAM will my results generate faster than on my local?

Mitch

Keep getting this error mat1 and mat2 shapes cannot be multiplied (77x768 and 4096x5120) Is this to do with width and height? I am matching the video info numbers to the blue nodes above the ref image.. what am i doing wrong?

Gabbis85

help plz, i got this i tryed it and it did not work, it say that it is installed but does not work still: Fast download using 'hf_transfer' is enabled (HF_HUB_ENABLE_HF_TRANSFER=1) but 'hf_transfer' package is not available in your environment. Try `pip install hf_transfer`. can this be turned off or do i have to use it? btw i did try to install it here: did type cmd in this folder "ComfyUI_windows_portable" then enter but did says it is installed there, fix was to edit the file: run_nvidia_gpu.bat then type this instead then it works: set HF_HUB_ENABLE_HF_TRANSFER=0 .\python_embeded\python.exe -s ComfyUI\main.py --windows-standalone-build pause

Cristi Tanase

got the same problem, 12GB VRAM, but I'm using the low memory model and still the WAN21_Vance seems not to load in video memory and everything works extremely slow

Gabbis85

help plz: DWPose from comfyui_controlnet_aux) does not use GPU, it render movie extremly slow atm, 2 sec clip takes 3h or so to make, i did fix the "DWPose from comfyui_controlnet_aux" part typing cmd in the install folder and then .\python_embeded\Scripts\pip.exe install onnxruntime-gpu and hit enter, but it still takes around the same time as before

Artokun

77 is not divisible by 2. you'll need to make sure all your dimensions are divisible by 2

EE ZC

where do your source your videos / clips , gifs ? its very slow when I use videos I make even smaller than 10 seconds im using RUNPOD on 4090

Mitch

but the numbers in the error are not the numbers I am using. The numbers 77x768 and 4096x5120 are not even in my workflow anywhere. I double checked and all my numbers are divisible by 2.

I Bea

When I run the workflow, I do not get an video output. Instead I have this msg. Exception in callback _ProactorBasePipeTransport._call_connection_lost(None) handle: Traceback (most recent call last): File "asyncio\events.py", line 88, in _run File "asyncio\proactor_events.py", line 165, in _call_connection_lost ConnectionResetError: [WinError 10054] An existing connection was forcibly closed by the remote host Can anyone help point me in the right direction?

Bio

I ran the autobat and chose the lesser ram option. But I can only chose "umt5_xxl_fp8_e4m3fn_scaled.safetensors" under "Load CLIP". Shouldn't I get the "T5_Q3_K_L"-version when chosing less than 16 gb ram? (3080 10 gb)

H-BOX

Hi ! This one is incredible. Could you just tell us what is the tweak to make follow more or less the video reference fidelity ?

FPV Juice

I got these error when it comes too the ksampler: The size of tensor a (66330) must match the size of tensor b (67320) at non-singleton dimension 1 Anybody else with this issue?

FPV Juice

I have these message... The size of tensor a (66330) must match the size of tensor b (67320) at non-singleton dimension 1 , no solution yet

Tim Eilers

In the future, can you PLEASE include the assets you used in the example workflow to aid with debugging in case we have issues? Thank you

GomezBro

yes obviously

Aitrepreneur

what resolution did you use for the generation?

Aitrepreneur

you could increase the lora causvid strength but it all depends on the video

Aitrepreneur

that's the text encoder, not the Vace model

Aitrepreneur

usually that error is not a big problem, rerunning the generation should work, otherwise the issue might be somewhere else

Aitrepreneur

I use something like giphy or pexels for the videos

Aitrepreneur

Have you tried to use the installer to reinstall comfyui from scratch?

Aitrepreneur

what resolution are you using for the generation?

Aitrepreneur

Why haven't you send me a dm like I asked if it's not working for you????

Aitrepreneur

might be a bad depth anything download, go inside that folder for you C:\\Users\\DDD\\Downloads\\ComfyUI_windows_portable\\ComfyUI\\custom_nodes\\comfyui_controlnet_aux\\ckpts\\depth-anything\\ and delete the Depth-Anything-V2-Large folder and try again

Mr Smith

honestly ive been a patreon for a while and i like the ease of the one click installers - but this is like the 3rd thing ive tried where there is not instructions on why things are broken or why things a so slow, the videos are becoming almost clickbait style - its not a good look man

Aitrepreneur

Just dm me your issues man how can i help fix hte issue if you don't tell me, all systems are different...

Herbert

i used 720x408 as input and output, same problem: KSampler The size of tensor a (57375) must match the size of tensor b (59670) at non-singleton dimension 1

Mr Smith

ok i have done

Gabbis85

hi :) what installer are you talking about? i have reinstalled it all with VACE-COMFYUI-MANAGER_AUTO_INSTALL.bat same issue sadly

Herbert

also my RAM runs nearly full during the KSampler (over 35GB usage)

Aitrepreneur

dm me your screenshot of the parameters before the generation

Mr Smith

its 100% using the CPU offload - i have 0 idea why I use the Q5 one cause i have 16VRAM and it took 47 mins to generate a 3 sec clip I have DM'd we'll see what happens

Aitrepreneur

it's probably because you are using a weird resolultion, for it to works it usually must be a multiple of 64, but you can use 480 and 720 numbers as well.

Alpha Ghost 47

PytorchStreamReader failed reading zip archive: failed finding central directory - Keep getting this error when I run a generation

Mr Smith

has anyone who is running this locally managed to get this running at the speeds shown in the video?

Gabbis85

0 clips done here, its takes forever for me, bad stats on the the task manager when rendering: cpu 17% memory 96% gpu 7%

Mr Smith

yeah its not using the GPU at all, im wondering if it is for anyone else? - my only one took that worked took 47 mins for a 3 secs clip on Q5 16vRAM GPU- its using CPU only, im curious if only the super RunPod gpus can handle it and therefore isnt really capable locally as advertised

Bio

Thanks for the answer. Since I read that you should use Q4_K_S + T5_Q3_K_L I got a little confused. All good now, works as it should.

H-BOX

Oh so that's the one handling either it follows the video of not ? ok !

Raziel Malichor

I've been meaning to ask. Does anyone know how to make LoRA's for these types of AI, I assume a video LoRA? And if our AI Overlord sees this, do you have any plans on making a video showing how to make video LoRA's, or are there no "easy" methods yet?

Dmitry Papyrin

Hi all, Aitrepereneur, i get the following issue. When running on my PC (4080) on Q5 model the proccess freezes on 87%, when it has to come to generation in WAN1, the log shows the following Attempting to release mmap (726) 0%| | 0/12 [00:00

Bio

I had the same, worked after "update all" in the ComfyUI manager.

Bob Winberry

Any idea what this error is about? I used your Auto Install and haven't changed anything: LoaderGGUF: cannot reshape array... thanks

BigBullAI

I started the Comfy UI in Runpod, but the Comfy UI version is the old one and I get an error when I want to update ComfyUI inside Comfy through the Manager. The workflow will not run. Hitting "Queue Prompt" does nothing. Anybody know how I can update ComfyUI in Rinpod/JupyterLab? Thanks.

Dmitry Papyrin

That helped, thank you. No when it comes to WAN it shows an error "torch.OutOfMemoryError: Allocation on device" and stops. Probably need to drop to lower model... or is it possible to fix? UPD: somehow it fixed with a restart. Though - the results of anymation are a bit not realistic but more of drawing style, has anybody fixed it?

David

....where is VACE_IMAGE-TO-VIDEO_CONTROLNET.json?

Russ Ader

Runs okay for me but takes like a half hour to generate a 81 frame video on an RTX 4070 TI Super (16 gigs vram). Also curious, if frame pack can do 2 minutes, how long can you generate a video for with this?

Josh T

Hey man, after installing everything with all nodes and packages, it's giving me this error when i try to run: Prompt execution failed Prompt has no outputs Any ideas?

Pizza

Unfortunately I can run the first bat file as it says it's 16bit and won't run on x64.

Junior

Had the same issue with bat file the first thing it wants to do is install git. Solved it by going to git website directly and downloaded the 64bit version and it worked.

Pizza

Ended up self installing it. 3 hours later I got my first rendering. It works! Now to play around with it, thanks.

Anthony J

i got this error when running: "* LoaderGGUF 21: - Value not in list: gguf_name: 'Wan2.1-VACE-14B-Q4_K_S.gguf' not in ['Wan2.1-VACE-14B-Q8_0.gguf']" Other parts seem to work.. i'm fairly new to comfyui but experienced in automatic11111.. there are some other errors but i think the incorrect vace name is causing the others. full out: got prompt Failed to validate prompt for output 8: * LoaderGGUF 21: - Value not in list: gguf_name: 'Wan2.1-VACE-14B-Q4_K_S.gguf' not in ['Wan2.1-VACE-14B-Q8_0.gguf'] Output will be ignored Failed to validate prompt for output 95: * VHS_LoadVideo 91: - Custom validation failed for node: video - Invalid video file: AnimateDiff_00079.mp4 Output will be ignored Failed to validate prompt for output 87: * VHS_LoadVideo 90: - Custom validation failed for node: video - Invalid video file: vaceupscale_00002.mp4 Output will be ignored WARNING: [WinError 2] The system cannot find the file specified: 'C:\\repo\\ext_repos\\ComfyUI_windows_portable\\ComfyUI\\input\\AnimateDiff_00079.mp4' WARNING: [WinError 2] The system cannot find the file specified: 'C:\\repo\\ext_repos\\ComfyUI_windows_portable\\ComfyUI\\input\\vaceupscale_00002.mp4'

Tim Eilers

Yes please, it helps when the first time it runs everything works, and then you know that (for example) using images that aren't 8x multiples resolution on each side will cause it to crash

Anthony J

ok , this was my fault cause i didnt exactly start from a clean slate. i first cloned all the git repos into a folder and started manually going through your guide, then got the patreon and ran it in a folder with a portable comfyui partially setup already. i found the error when i did a comfymanaager update, it had some error about permissions on the git folder. fixing that and updating the nodes fixed the output

Jon Bott

I see scripts for RUNPOD and for Windows. Is there anything for MacOS?

troy X

The post below these comments

troy X

Ran fine and rendered fine, but after I close ComfyUI. How do I reopen. Where did it install/extract to?

Timothy Musson

Dmitry, what happens if you disable the Wan21_CausVid LoRA (via the workflow's Power Lora Loader node)? For me, that solves the "drawing style" look. But either way, the processing time is slow. (Basically, the CausVid LoRA doesn't seem to be working correctly - at least in my case.)

Dmitry Papyrin

Jut tried that, but it completly messed the result - the result was very abstract, does not bring the real look on the refference image, and the dots and lines from video combine nod appeared on the image. I guess what i need is to play with Lora strengeth and KSampler settings.

Junior

Can you do a tutorial vid on the vace install from scratch locally not the run pod? This will save a lot of time and cut down on the barrage of emails you are probably getting.

Tommi

I tried it both local and runpod both censored to me

PLLPower

Make sure to enable the part of the workflow you want to use. On the left of the window, you should have another much smaller window that you can use to enable it.

Junior

How did you install

Vaivij

Is there a way to mask just a portion of the video? I want to just change the name of the book the person is holding in my video. Or a prompting feature that does that? It currently changes the whole person and makes him look animated.

ПЕТЪР ЕВТИМОВ

Can all your AI workflows work with NVIDIA RTX 5090? Thanks for all your great work!

Aitrepreneur

Actually runpod is slower than local gen, but at the same time I can only speak because of my 4090

Aitrepreneur

For this you need to train a lora for wan, might do a video about it if people are interested

Aitrepreneur

everything depends on the prompt and source image + video, causvid lora has nothing to do with it

Aitrepreneur

use the Q4 model instead maybe for faster gen, it's probably loading partially for this to be that slow. The limit is not seconds but the number of frames, the more frames there are, the les coherent the videos start to be become (around 120 frames is where things starts to be bad)

Aitrepreneur

I made a document for the manual install, it's the same thing all the time.

Aitrepreneur

"CENSORED"??? This is a controlnet my man...nothing censored here, you can input anything you want as a base video and reference image....

Aitrepreneur

Technically the vace model is capable of that but the functionality is not implemented yet inside comfyui to work correctly, wait and see until it does.3

Aitrepreneur

Workflows are not the issue, it's just a workflow, what you need is use the latest version of comfyui for that. So yes.

Dmitry Papyrin

thank you, for some reason.. I'm stuck again at the same spot. will try to use lower version...

Dmitry Papyrin

I've noticed that LoadClip node uses fp8 safetensors, not q5, can this be a reason?

Tommi

do I need to use Loras?

Aitrepreneur

for the vace installer and comfyui???? absolutely not

Raziel Malichor

I sincerely hope that you do make a video about it. Your videos make this stuff so easy to understand and I wouldn't have a clue how to do it otherwise xD Hopefully other people are interested too. Keep up the incredible work!

Ruiz Lihuel Alejandro

Thanks for the quick response. I'm updating the nodes. Maybe I shouldn't do it then?

Timothy Musson

Thanks Aitrepreneur, I finally got it working - it's just a squeeze on 12Gb cards :) Here's how I got it working on my 12Gb 3060, in case it helps anyone else... The 14B GGUF VACE models and the text encoder model together total over 12Gb, even when using the Q4 version of the VACE model. (The workflow outputs video, but it's super slow, and the video looks rather contrast-y and weird.) Turns out, the 1.3B VACE safetensors model can be used instead: at 4.3Gb, it's small enough to fit alongside the text encoder. You can find it here, named wan2.1_vace_1.3B_fp16.safetensors: https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged/tree/main/split_files/diffusion_models Place that file in ComfyUI/models/unet/ In the workflow, replace the "GGUF Loader" node with a "Load Diffusion Model" node. In that "Load Diffusion Model" node, select the wan2.1_vace_1.3B_fp16.safetensors model. I'm sure there's a quality hit, but hey, it works pretty nicely anyway. With this I'm generating 81 frame animations at 432 x 768 in just over 3 minutes. And... here's a hacky sort of tip, I guess: To get slightly longer videos, use a video editor (e.g. kdenlive for me on GNU/Linux) to double the speed of the input video. Then use the workflow's interpolation group to smoothly "slow" the result back down to a sensible speed. Works pretty well :)

Pak26

The result is always an oversaturated mix between the image and the video sources . Is there any ways to preserve the original image and only have the movement from the controlnet applied to it ? Thanks

BigBullAI

Hi Aitrepeneur, thanks for this. First I’d like to point out a few minor mistakes in your Runpod instructions, which yu might want to fix: In 4., I think it should say "Connect to Jupyter Lab port 8888" instead of "Connect to HTTP SERVICE port 8888". In 5., it should say VACE_AUTO_INSTALL-RUNPOD.sh instead of HIDREAM_AUTO_INSTALL-RUNPOD.sh Here’s my question: everything runs now, but I don’t get an output. Why?

Gabbis85

thx so much worked for me! :) was getting this massege tho when rendering: ERROR lora diffusion_model.time_projection.1.weight shape '[9216, 1536]' is invalid for input of size 157286400 WARNING SHAPE MISMATCH diffusion_model.time_projection.1.bias WEIGHT NOT MERGED torch.Size([30720]) != torch.Size([9216]) i think LoRA .safetensors not 100% compatible

Mr Smith

hey man could you tell me step by step how to add a Load Diffusion model and i dont see it, thanks

BigBullAI

After I updated everything again, it worked. Another question: How can I control to take more from the reference image and less from the video?

Gabbis85

hi, Right click bakground in comfyui then "add node" then "advanced" then "loaders" then "load Diffusion Model" then you have to connect the cable the one from the other model i mean from the "gguf loader" but perhaps dont have to use this as later i was able to use: Wan2.1-VACE-14B-Q4_K_S.gguf fix was to change the width & height & fps & frame for every video Also sett KSampler "steps" to 4 instead of 12 perhaps, as shown in the tutorial, i dont use wan2.1_vace_1.3B_fp16.safetensors

Dmitry Papyrin

I did some research and what appeared is the issue with VRAM. In short: What i considered previously a freeze on KSampler is not a freeze, but a super slow processing. Depending on resolution of output the VRAM and size of model VRAM is overloaded (My 16GB 4080). Traditionally, if VRAM ran out, the process would crash. NVIDIA introduced a new driver feature (since driver 536.40) that tries to automatically offload overflow GPU memory to system RAM instead of crashing. This is called “CUDA - System Fallback.” It was intended to let SD models continue running on low-VRAM GPUs (at a much slower speed) rather than erroring out On driver 546.01 and above, NVIDIA added an option to disable this behavior. If disabled there would be no freeze, but a crash. 1.3B is much lighter model, so it helps to fix this issue. I tried you solution – to switch to 1.3B and it fixed the issue for me, but still I’m not able to get the proper result. Now the result looks almost like an outcome from Video Combine nod. What did you use on weight_dtype of Load Diffusion Model nod? UPD: checking Loras for 1.3B

Pete

Hello, I'm using a rtx 3090 but its taking 30mins+ to generate anything. Is this normal?

troy X

Nevermind It was in my downloads folder but hidden. You have to type it. Took me sometimes to find it. I used AItrepeneur's auto install batch files.

montag bradbury

Just joined. Thanks. This is working quite well for my 4070 GPU. I think that much of the problem people are having is due to lack of familiarity with ComfyUI and some its particular quirks. You might want to let people know that even if they have a 12GB VRAM GPU they should probably install below that, or it will tend to crap out. Also, if you get an error msg I have found that if you just click the run button again it will TEND to continue from where it bombed. Most important though is installing with the lower VRAM model.

Jaybli

Got the same problem, any1 find a solution? Tried 1080 x 720, and nothing. Same error.

Timothy Musson

Dmitry, I left 'weight_dtype' on 'default', so I'm not sure which one it's picking. (Come to think of it, the options on my system are: 1) default, 2) fp8_e4m3fn, 3) fp8_e4m3fn_fast, and 4) fp8_e5m2. Yet the 1.3B model is labelled fp16. So, I'm guessing none of the options are ideal.)

성운 이

I want to use inpainting in the video. Please tell me how

Mazena

Whenever I try to do an image to video generation, it gets stuck here with 0%: Requested to load WanVAE loaded completely 11927.319667816162 242.02829551696777 True Requested to load WAN21_Vace loaded partially 14994.802950408935 14994.797119140625 0 Attempting to release mmap (337) 0%| | 0/12 [00:00

robin poirieux

Thanks for your video and workflow. It works perfectly, although rendering is still relatively slow with the LoRA: 30 minutes using the smallest model on an RTX 4060 Ti 16GB. Could you explain how to add more LoRAs? The node does offer the option, but I assume the standard LoRAs need to be converted to a video-compatible version, right?

Epic Cardboard Props (edited)

Comment edits

2026-09-08 20:00 Hi all, something weird is happening in this workflow for me, I always end up with the ControlNET colored lines and white dots in the final videos, the only way I don't get that issue is when I enable the Lora which makes the final render much less realistic. Anybody else with this issue?
2025-05-26 19:40

Aitrepreneur

you can simply decrease the "denoise" value in the ksampler node to something like 0.93 for less contrasted videos

Aitrepreneur

you can play with the strength value and maybe decrease the denoise value as well to 0.93 for example

Aitrepreneur

try downloading and using the Q4 model instead

Aitrepreneur

you're getting out of vram, try downloading and using the Q4 model instead

Aitrepreneur

I'll be making more video worfklows, I'll put it on the list

Aitrepreneur

try downloading and using the Q4 model instead, it will be faster for you. For the loras you need to use special wan loras and you can load them via the power lora loader

theworldofanatnom

is there anyway around the fact that if i use a lora then the generated image is more like the training data than the input image

Pak26

I forgot to mention an option for I2V would also be nice ! Thanks again ! :)

성운 이

I want to make the video time longer while continuing to expand the end of the video. Is there a way?

BirbFood

Which generator settings most control speed of the generation? KSampler Steps, per your video, but what else? WanVaceToVideo Length of frames? The Width and Height? The number of Muters active? Does anything make it much slower aside from the steps? Thank you!

Gabbis85

To clarify i did get an error starting the rendering so i did turn this error off: DepthAnythingV2Preprocessor Fast download using 'hf_transfer' is enabled (HF_HUB_ENABLE_HF_TRANSFER=1) but 'hf_transfer' package is not available in your environment. Try `pip install hf_transfer`. fix was to edit the file: "run_nvidia_gpu.bat" to this: set HF_HUB_ENABLE_HF_TRANSFER=0 .\python_embeded\python.exe -s ComfyUI\main.py --windows-standalone-build pause to clarify i did turn this off and it worked fine i think, can atleast make clips, but do i really need HF_TRANSFER? anyone else has this off? i did ofc try manually install hf_transfer, but did not work for me

Nicolas Diaz

Hey there just wanted to point out there was likely an error in your runpod instructions. You have us dragging over the HiDream.bat file then inputting into terminal the run command for VACE. I am a newbie so I somehow thought HiDream needed to be installed to run VACE because of this.

Yo

When I try to auto-install "VACE-COMFYUI-MANAGER_AUTO_INSTALL.bat", an error message saying "Unsupported 16-Bit Application" shows up and it shuts down. Here is my system info: Processor Intel(R) Core(TM) i9-14900HX 2.20 GHz Installed RAM 32.0 GB (31.7 GB usable) System type 64-bit operating system, x64-based processor

relaxama

* LoaderGGUF 21: - Value not in list: gguf_name: 'Wan2.1-VACE-14B-Q4_K_S.gguf' not in ['Wan2.1-VACE-14B-Q8_0.gguf'] Output will be ignored Prompt executed in 3.44 seconds got prompt Failed to validate prompt for output 8: * LoaderGGUF 21: - Value not in list: gguf_name: 'Wan2.1-VACE-14B-Q4_K_S.gguf' not in ['Wan2.1-VACE-14B-Q8_0.gguf'] Output will be ignored Prompt executed in 0.06 seconds if its liek this do i run again the model noeds install bat?

Aitrepreneur

you need to use to lora for the fast generation, or else it will not work

Aitrepreneur

this means that you are not selecting your downloaded model in the first node

relaxama

got it . thanks man you are the best.

Kittycha and Waifus (edited)

Comment edits

2026-09-08 20:00 Hello, two custom nodes are not appearing in ComfyUI, namely WanVaceToVideo and TrimVideoLaten
2025-05-29 17:29

Russ Ader

I have no idea how you are getting this to work in a couple minutes, even with a RTX 4090 and the larger model. I have downgraded all the way down to the Q3 version and it is still taking me over 20 minutes to generate a few seconds. I have an RTX 4070 TI Super (16 gigs vram) and this is slow as snot. Even with the Q3 model my vram and GPU are completely maxed out. EDIT: Just looked again, it's at 20 minutes and says it will take another 10 minutes to finish. Thats over 30 minutes with a Q3 model.

Aitrepreneur

update comfyui correctly and relaunch the workflow

Aitrepreneur

check your resolution and number of frames, I use 720x720 resolution max with 81 frames

Fredden

Hi! I get this error: The size of tensor a (15276) must match the size of tensor b (15504) at non-singleton dimension 1

Aitrepreneur

use a resolution that a multiple of 8/64. I like using 720x720 for a square and 720x480 for vertical/horizontal

simon myers

I get this error when I try to make a video, same thing happens when any of the control nets are used. I've checked the file path and nothing is there. I've updated comfy ui and checked for missing nodes, but it shows nothing is missing. Any ideas? edit- I can run text and image to video wan just fine. I installed using your installer - VACE-COMFYUI-MANAGER_AUTO_INSTALL DWPreprocessor [Errno 2] No such file or directory: 'C:\\Users\\xxxx\\OneDrive\\Documents\\_PROJECTS\\_AI_ART\\COMFYUI\\ComfyUI_windows_portable\\ComfyUI\\custom_nodes\\comfyui_controlnet_aux\\ckpts\\yzd-v\\DWPose\\.cache\\huggingface\\download\\0XR-wYEaL4qLqwIO4oYox_j1wmI=.7860ae79de6c89a3c1eb72ae9a2756c0ccfbe04b7791bb5880afabd97855a411.incomplete'

unreal sport

when i put an anime scene as a reference image in the video production i lose the style. Is it possible to keep the style of the reference image intact?

Mr Smith

this is exactly what happened with me - Q5 - 47 mins - q4 - 22 mins Theres defo some clickbait goin on - i just got told that something else must be hogging the VRAM and basically - sorry i got a better card than you - i have a 16vram 4060Ti and theres 0 chance this is being able to run in a few mins

Vicente Varela

Hi, I keep getting this message:DepthAnythingV2Preprocessor Input type (torch.cuda.FloatTensor) and weight type (torch.FloatTensor) should be the same

Russ Ader

I found if I drop to 480 it takes about 3 to 4 minutes per second with a 4070 ti super (16 gigs vram)

Russ Ader

Okay, I have played around with this and Framepack Studio almost nonstop for the last two days to figure them out and here are my conclusions, which may help someone and if anyone is willing to take up my proposal for a workflow and succeeds at doing it, they will blow the video ai world away. First a comparison of the two: For Vace: Anything less than a RTX 4090 will do better with 480p no matter how good your card is. I consider mine, RTX 4070 TI Super, to be the bare minimum to pull off 720p and it takes around 10-12 minutes to per second. At 480 it takes about 3 minutes per second with the Q5. Decreasing the frames per second and increasing the number of times I perform interpolation, meaning running it over and over, not increasing the number of frames in one pass, will usually reduce the amount of blurring and flickers. Framepack studio: Gets funky after about 5-10 seconds. Watching the process I realized that what is actually happening is that it is generating 2-3 second clips at a time and interlacing them. It is creating the beginning or end frame, depending on if you are using the original or the F1 version, and connecting the last image it makes to the previous. This is why it continues to distort the longer the video, because it moves further from the original image as time goes on. Framepack studio also is missing one huge factor the the original Framepack installer on this patreon has, the ability to choose your compression rate. Compression should only be done after the last process so that it can produce cleaner, better detailed videos. So any workflows moving forwards should have a 0 compression rate until the end of all workflows where it is an option to turn on from being the the default off, that way if you need to cycle the video back through again you can. However, the new Framepack studio has an interesting concept, to inject a start and end frame. The best way to solve the video generation issue is to create a workflow that allows a person to upload an image multiple times throughout the process. So you start with the beginning image, then set an end image. Then that end image is the beginning image for the next 3 seconds and so on. By hopping from one image to the other you will be preventing the very thing that is holding both of these up, distortion. Framepack is likely to work better when it comes to this idea, which is a shame because Vace gives better control with controlnets.

Aitrepreneur

yes that's why you use an anime image for the reference

Taiga

D:\VID-AI\ComfyUI\input\20250522_1122_Cosmic Photographer Adventure_simple_compose_01jvvhcx5yexzbmk5kytp9eg0b.mp4 could not be loaded with cv.

Taylor Sheahan and The Dusty Desert Boys

I keep getting this K sampler error: The size of tensor a (66330) must match the size of tensor b (67320) at non-singleton dimension 1 here is the log: # ComfyUI Error Report ## Error Details - **Node ID:** 2 - **Node Type:** KSampler - **Exception Type:** RuntimeError - **Exception Message:** The size of tensor a (66330) must match the size of tensor b (67320) at non-singleton dimension 1 ## Stack Trace ``` File "D:\vanc\ComfyUI_windows_portable\ComfyUI\execution.py", line 349, in execute output_data, output_ui, has_subgraph = get_output_data(obj, input_data_all, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "D:\vanc\ComfyUI_windows_portable\ComfyUI\execution.py", line 224, in get_output_data return_values = _map_node_over_list(obj, input_data_all, obj.FUNCTION, allow_interrupt=True, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "D:\vanc\ComfyUI_windows_portable\ComfyUI\execution.py", line 196, in _map_node_over_list process_inputs(input_dict, i) File "D:\vanc\ComfyUI_windows_portable\ComfyUI\execution.py", line 185, in process_inputs results.append(getattr(obj, func)(**inputs)) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "D:\vanc\ComfyUI_windows_portable\ComfyUI\nodes.py", line 1525, in sample return common_ksampler(model, seed, steps, cfg, sampler_name, scheduler, positive, negative, latent_image, denoise=denoise) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "D:\vanc\ComfyUI_windows_portable\ComfyUI\nodes.py", line 1492, in common_ksampler samples = comfy.sample.sample(model, noise, steps, cfg, sampler_name, scheduler, positive, negative, latent_image, ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "D:\vanc\ComfyUI_windows_portable\ComfyUI\comfy\sample.py", line 45, in sample samples = sampler.sample(noise, positive, negative, cfg=cfg, latent_image=latent_image, start_step=start_step, last_step=last_step, force_full_denoise=force_full_denoise, denoise_mask=noise_mask, sigmas=sigmas, callback=callback, disable_pbar=disable_pbar, seed=seed) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "D:\vanc\ComfyUI_windows_portable\ComfyUI\comfy\samplers.py", line 1133, in sample return sample(self.model, noise, positive, negative, cfg, self.device, sampler, sigmas, self.model_options, latent_image=latent_image, denoise_mask=denoise_mask, callback=callback, disable_pbar=disable_pbar, seed=seed) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "D:\vanc\ComfyUI_windows_portable\ComfyUI\comfy\samplers.py", line 1023, in sample return cfg_guider.sample(noise, latent_image, sampler, sigmas, denoise_mask, callback, disable_pbar, seed) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "D:\vanc\ComfyUI_windows_portable\ComfyUI\comfy\samplers.py", line 1008, in sample

Sam Mesri

getting this error: easy imageRemBg CUDA error: no kernel image is available for execution on the device CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect. For debugging consider passing CUDA_LAUNCH_BLOCKING=1 Compile with `TORCH_USE_CUDA_DSA` to enable device-side assertions.

VXLM

can you give the solution here? cuz i got the same thing. running on runpod

Fang

Please help, I have tried every method of installation listed above and still can not get the pose and depth controlnet to work. I keep getting these two errors. The canny controlnet works fine. DWPreprocessor [Errno 2] No such file or directory: 'D:\\Program Files\\ComfyUI_windows_portable_nvidia\\ComfyUI_windows_portable\\ComfyUI\\custom_nodes\\comfyui_controlnet_aux\\ckpts\\yzd-v\\DWPose\\.cache\\huggingface\\download\\0XR-wYEaL4qLqwIO4oYox_j1wmI=.7860ae79de6c89a3c1eb72ae9a2756c0ccfbe04b7791bb5880afabd97855a411.incomplete' DepthAnythingV2Preprocessor [Errno 2] No such file or directory: 'D:\\Program Files\\ComfyUI_windows_portable_nvidia\\ComfyUI_windows_portable\\ComfyUI\\custom_nodes\\comfyui_controlnet_aux\\ckpts\\depth-anything\\Depth-Anything-V2-Large\\.cache\\huggingface\\download\\EheDQYkKfRTio82VDltlF5Af82M=.a7ea19fa0ed99244e67b624c72b8580b7e9553043245905be58796a608eb9345.incomplete' I do get this error though when installing the vace files: 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. mediapipe 0.10.21 requires numpy<2, but you have numpy 2.2.6 which is incompatible. mediapipe 0.10.21 requires protobuf<5,>=4.25.3, but you have protobuf 6.31.1 which is incompatible.

Josh inmn

WHAT DOES GEMINI 2.5 PRO SAME WHEN YOU PASTE THE ERROR IN THERE?

Josh inmn

WHAT DOES CLAUDE 3.7 SUGGEST YOU DO WHEN ASK?

F C

hi love your creations, i'm getting this error when trying to do video to image :LoaderGGUF cannot reshape array of size 13834627 into shape (5120,2880) comfyui

James Stewart

Hi folks, thoughts on this error running the VACE Comfy auto installer: File "C:\Users\Username\Downloads\ComfyUI_windows_portable\python_embeded\Lib\site-packages\transformers\utils\versions.py", line 44, in _compare_versions raise ImportError( ImportError: huggingface-hub>=0.30.0,<1.0 is required for a normal functioning of this module, but found huggingface-hub==0.28.0. Try: `pip install transformers -U` or `pip install -e '.[dev]'` if you're working with git main Press any key to continue . . . I'm guessing I need to run this via CMD, just not sure on how to get to the right path? (not as techie as some of you)

JMC3DX

I dont get it @Aitrepreneur Like Klink on the (payed) websites it maintains character consistency. With this work flow i get a different face then in the source image. Any way to keep it more constant / the same as the face in the source image?

Ruiz Lihuel Alejandro

Same problem here. This answer from Aitrepreneur worked for me: use a resolution that a multiple of 8/64. I like using 720x720 for a square and 720x480 for vertical/horizontal

Davi Carvalho

KSampler The size of tensor a (83600) must match the size of tensor b (87400) at non-singleton dimension 1

lurkergao

it says your dependency is out of date. first try going to the manager button and hitting update all. reload workflow see if that worked. sometimes some of the downloads fail when running the auto installer. try running it again.

Andrew Saltmer

Ran the Vace-Models-Nodes-Install.bat in my existing ComfyUI installation and ComfyUI would not work, would load in the console then close. So installed a clean installation to my desktop, checked that worked, then dragged the new install into my existing ComfyUI location and it now works.

Fang

I am also having this issue. The generated faces don't even remotely match the source image. Any advice would be appreciated

rc

Got same error. Solve it by using a reference square image of 1024x1024. Many of the out of sight overlapping node use square ratio. Output set to the video ratio.

Anthony

New PC, windows 11. I get a error a The program or feature "\??\E:\git-setup.exe" cannot start or run due to incompatibity with 64-bit versions of Windows. Please contact the software vendor to ask if a 64-bit Windows compatible version is available.

Martin Marković

I assume this does not create virtual environment and can fuk up my current python ?

Gimmy Grayn

Got same error on Win10 64-bit

Dustin Lausch

To be safe I always use Conda to create an ENV and then activate it, run the .bat files from inside the ENV and it seems to keep things safe outside that shell.

Anthony

Now installed 😀

James Stewart

Thanks @George Thompson - Ok I've done all that, reinstalled and updated the manager etc. I've tried generating an image in Comfy (worked before) but now getting a "torch.OutOfMemoryError: Allocation on device" message I know @Aitrepreneur mustn't be well but I'm increasingly at a loss and really want to get things working! I'm not super techie and bumble along usually successfully, but this time I feel I'm stuck in a loop. I get the feeling I'm going to have to delete everything from Comfy and start totally again...

Steve

yea this puts out something that is worse quality than the original video and combines that with something that only remotely resembles the source image. Bummed that it was hyped so much.

Steve

a lot of this is accurate...that first paragraph though... I would say with 4070 its closer to 100 seconds per second of video using q5 ...but that would also depend on the video's FPS. Maybe I'm using less and you're using more.

Gimmy Grayn

Hi, launched your workflow on RunPod, did everything according to the instructions from the video on youtube. chose 4090 video card, launched everything, updated and launched comfyui. After that everything started, but when generating a video I get an error - KSampler The size of tensor a (19080) must match the size of tensor b (19440) at non-singleton dimension 1 On my home computer with 3080ti card 12gb i got the same error, only the number of tensor was different than on runpod. I conclude that your build has some kind of problem that manifests itself both on the runpod and on my PC. Сan you help fix this?

Gimmy Grayn

got prompt 0%| | 0/12 [00:00 lambda input, sigma, **kwargs: predict_eps_sigma(model, input, sigma, **kwargs), ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/workspace/ComfyUI/comfy/extra_samplers/uni_pc.py", line 843, in predict_eps_sigma return (input - model(input, sigma_in, **kwargs)) / sigma ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/workspace/ComfyUI/comfy/samplers.py", line 396, in __call__ out = self.inner_model(x, sigma, model_options=model_options, seed=seed) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/workspace/ComfyUI/comfy/samplers.py", line 945, in __call__ return self.predict_noise(*args, **kwargs) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/workspace/ComfyUI/comfy/samplers.py", line 948, in predict_noise return sampling_function(self.inner_model, x, timestep, self.conds.get("negative", None), self.conds.get("positive", None), self.cfg, model_options=model_options, seed=seed) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/workspace/ComfyUI/comfy/samplers.py", line 376, in sampling_function out = calc_cond_batch(model, conds, x, timestep, model_options) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/workspace/ComfyUI/comfy/samplers.py", line 206, in calc_cond_batch return executor.execute(model, conds, x_in, timestep, model_options) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/workspace/ComfyUI/comfy/patcher_extension.py", line 111, in execute return self.original(*args, **kwargs) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/workspace/ComfyUI/comfy/samplers.py", line 325, in _calc_cond_batch output = model.apply_model(input_x, timestep_, **c).chunk(batch_chunks) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/workspace/ComfyUI/comfy/model_base.py", line 151, in apply_model return comfy.patcher_extension.WrapperExecutor.new_class_executor( ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/workspace/ComfyUI/comfy/patcher_extension.py", line 111, in execute return self.original(*args, **kwargs) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/workspace/ComfyUI/comfy/model_base.py", line 189, in _apply_model model_output = self.diffusion_model(xc, t, context=context, control=control, transformer_options=transformer_options, **extra_conds).float() ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/local/lib/python3.11/dist-packages/torch/nn/modules/module.py", line 1553, in _wrapped_call_impl return self._call_impl(*args, **kwargs) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/local/lib/python3.11/dist-packages/torch/nn/modules/module.py", line 1562, in _call_impl return forward_call(*args, **kwargs) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/workspace/ComfyUI/comfy/ldm/wan/model.py", line 563, in forward return self.forward_orig(x, timestep, context, clip_fea=clip_fea, freqs=freqs, transformer_options=transformer_options, **kwargs)[:, :, :t, :h, :w] ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/workspace/ComfyUI/comfy/ldm/wan/model.py", line 696, in forward_orig c_skip, c[iii] = self.vace_blocks[ii](c[iii], x=x_orig, e=e0, freqs=freqs, context=context, context_img_len=context_img_len) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/local/lib/python3.11/dist-packages/torch/nn/modules/module.py", line 1553, in _wrapped_call_impl return self._call_impl(*args, **kwargs) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/local/lib/python3.11/dist-packages/torch/nn/modules/module.py", line 1562, in _call_impl return forward_call(*args, **kwargs) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/workspace/ComfyUI/comfy/ldm/wan/model.py", line 244, in forward c = self.before_proj(c) + x ~~~~~~~~~~~~~~~~~~~~^~~ RuntimeError: The size of tensor a (19080) must match the size of tensor b (19440) at non-singleton dimension 1 Prompt executed in 0.27 seconds

daniel baychev

Hello I am new to comfy and ai generations, but ive been studying how everything works the past couple of days and really experimenting with this workflow. If someone can have the time to explain some stuff to me i would really appreciate it. What i am trying to achieve is a character conaiatency (from the input image) when generating videos, but from my experience i always get a a totally new face and sometimes even the video reference gets ignored. I tryed to debunk why myself with gpt, but alas no success and thus im coming here for help 😄

Ruti Bindehassan Gayle

I need your help, overlord. I have followed the instructions and the console stops at "press any key to continue" they closes and won't launch.

jonnylpg1993

DWPreprocessor [Errno 2] No such file or directory: 'D:\\Program Files\\ComfyUI_windows_portable_nvidia\\ComfyUI_windows_portable\\ComfyUI\\custom_nodes\\comfyui_controlnet_aux\\ckpts\\yzd-v\\DWPose\\.cache\\huggingface\\download\\0XR-wYEaL4qLqwIO4oYox_j1wmI=.7860ae79de6c89a3c1eb72ae9a2756c0ccfbe04b7791bb5880afabd97855a411.incomplete' Does anyone have a solution or where to download and place the file to make it work?

jonnylpg1993

It fixed for my depth error but still getting the DWPreprocessor error. Is there another file im missing and where do I get it and place it? Im such a noob at this

Boy Thee

that's saying that the model did not download completely. so download the model manually and place it in \DWPose the other dirs are install trash

Boy Thee

the first error means the download did not finish. download the model manually, then drop it into Depth-Anything-V2-Large and delete the lower directories. there's a page that shows the correct paths.

Rick Magnun

Any updates coming soon???

Muriithi Nyaga

I am getting this error on my rtx 4060 16gb: DepthAnythingV2Preprocessor No operator found for `memory_efficient_attention_forward` with inputs: query : shape=(1, 2443, 16, 64) (torch.float32) key : shape=(1, 2443, 16, 64) (torch.float32) value : shape=(1, 2443, 16, 64) (torch.float32) attn_bias : p : 0.0 `fa3F@0.0.0` is not supported because: xFormers wasn't build with CUDA support requires device with capability > (9, 0) but your GPU has capability (8, 9) (too old) dtype=torch.float32 (supported: {torch.bfloat16, torch.float16}) operator wasn't built - see `python -m xformers.info` for more info `fa2F@0.0.0` is not supported because: xFormers wasn't build with CUDA support dtype=torch.float32 (supported: {torch.bfloat16, torch.float16}) operator wasn't built - see `python -m xformers.info` for more info `cutlassF-pt` is not supported because: xFormers wasn't build with CUDA support

halloking123

this means that the size of your A does not match the size of your B in this case something is wrong with size A

Muriithi Nyaga

I figured it out, I just replaced the depth anything V2 node with the regular Depth anything node

Michael O

I am getting: OutOfMemoryError, on a 8 GB VRAM, NVIDIA GeForce RTX 4060, is there anything i can do to lower it down?

Frank B

OT: Any chance of an All in one install and Comfy work flow for the newly released flux kontext? been using it in flux playground and its pretty damn amazing.

Jayson King

Can anyone give me suggestions for improving the faces of people at a mid to far distance within this workflow? i get good results if im close up like the demo but as soon as i run this on people mid to far, the faces are smudged. i was considering trying face swapping. ive upped the steps and even tried the cause vid v2... but im a noob so need some real expert advice.

theworldofanatnom

to anyone not able to get the nunchaku nodes: got to Comfyui\ComfyUI_windows_portable\update and update comfyui and python dependencies. the manager update button still doesn't work

Padraig Kane

i am getting this error i tried my best to download new models for small, base, large, and giant nothing seems to work can someone help me out please i placed the models in the \download\ folder for each i restarted it over and over and its not working Errno 2] No such file or directory: 'C:\\Users\\35385\\Downloads\\ComfyUI_windows_portable\\ComfyUI\\custom_nodes\\comfyui_controlnet_aux\\ckpts\\depth-anything\\Depth-Anything-V2-Small\\.cache\\huggingface\\download\\Wgzz53Hpl7-m749TyT1Y6_XeYfg=.715fade13be8f229f8a70cc02066f656f2423a59effd0579197bbf57860e1378.incomplete'

Padraig Kane

never mind i fixed it yay im learning if anyone as a problem like this place the here \ComfyUI_windows_portable\\ComfyUI\\custom_nodes\\comfyui_controlnet_aux\\ckpts\\depth-anything\\Depth-Anything-V2-Small all so had to do the same thing for DWpose

Padraig Kane

C:\Users\35385\Downloads\ComfyUI_windows_portable>pause. anyone know why it keeps pausing when i'm running vace. it pauses at 85% on the ksampler is this just a performance issue with my pc here are my specs : AMD Ryzen 5 5600G with Radeon Graphics 3.90 GHz ram hyperx 3200 16.0 GB (11.2 GB usable) RTX 3060 12GB if it is a performance issue is there any other image to video workflows for a lower end system? im also using the Q5 guuf model

Aitrepreneur

yeah it's your GPU, try using Q4 but Vace is really power hungry

Padraig Kane

Ok thanks for the help I just downloaded your hugging face link 🙂

Padraig Kane

it worked! but it the took 20 mins to execute the prompt ahahahah, what gpu would you recommend buying for video generation ill probably use runpod for the time being when generating vids. thanks for the help!!

Aitrepreneur

Vram is king so any gpu with as much vram as possible like maybe a second hand 3090 might be a good choice

Padraig Kane

i generated a few vids but they look cartoony compared to the pics a videos i used to generate any ideas how to get the outcome to look more realistic? @aitrepreneur

Aitrepreneur

I mean depends on some many things, resolution, prompt, base images, etc...

Timmy

I kept getting the error below for the VAE and the CLIPLoader: Error while deserializing header: MetadataIncompleteBuffer VAE: wan_2.1_vae.safetensors CLIPLoader: umt5_xxl_fp8_e4m3fn_scaled.safetensors I had to manually download the files from higgingface here: VAE: https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged/tree/main/split_files/vae CLIPLoader: https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged/tree/main/split_files/text_encoders Then copy them to their respective folders and overwrite the ones that were downloaded from the one click installer: VAE: \ComfyUI_windows_portable\ComfyUI\models\vae CLIPLoader: \ComfyUI_windows_portable\ComfyUI\models\text_encoders Also the ComfyUI Manager would not autoupdate using the UI. I had to manually use the "update_comfyui.bat" batch file, and it seemed to update ComfyUI to 0.3.44 at the time of this post. Hope this helps.

Timothy Danforth

How do I install a custom checkpoint to a ComfyUI instance on runpod? I've tried using the Jupyter UI to manually upload it to the workspace/ComfyUI/models/checkpoints directory, I've tried uploading to the base directory and using the terminal to move it, I've tried direct download from the terminal with wget, and I've even confirmed that the files are present in the directory, complete, and working... but nothing shows up in the workflow node dropdown except the default model from initial installation.

Whitecrow_69

Which model files did you download from hugging face? Was it the .pth files? Did you just place the .pth file in the depth-anything\\Depth-Anything-V2-Small folder? That’s still not working for me, any other pointers?

Padraig Kane

i think its just high saturation it makes it look cartoony

Aitrepreneur

Just send me a dm next time man, I can't see all the comments. Try downloading this file: https://we.tl/t-xImcxXUGO6 and put it in the C:\Users\steve\Downloads\ComfyUI_windows_portable\ComfyUI\custom_nodes\comfyui_controlnet_aux\ckpts\hr16\DWPose-TorchScript-BatchSize5 folder

Quân Vũ Trọng

"I’m having the same issue. The link you sent above is no longer working - could you please resend it to me?"

StatusGear

I got the error : 'SymmetricPatchifier' object has no attribute 'get_grid' . I assume it has to do with not having get_grid probably from some nodes. I do have comyfui_controlnet_aux (60 conflicts) , comfyUI-GGUF (2 conflicts) , and ComfyUI-KJNodes (2 conflicts). But I just want to reconfirm.

Aitrepreneur

send me some screenshots of the parameters before and after the generation

Madhy Odreman

Runing the COMFYUI for the first time and I am getting the following error: Cannot start or run due to compatibility with 64-bit version of Windows. Please contact the software vendor to ask if a 64-bit Windows compatible version is available.

Dr. Brian Burton

I am getting the error VHS_LoadVideo cannot allocate array memory. Any suggestions on where to start troubleshooting?

Aitrepreneur

that is a lack of vram, what are you parameters here?

Dr. Brian Burton

Thank you. Once I tweaked the vram settings, everything worked.

James Bowler

I have a RTX3090 with 24GB and I followed your video, in your video, you are using this gguf: Wan2.1-VACE-14B-Q4_K_S.gguf, When I downloaded the Q8 version, your workflow defaults to Wan2.1-VACE-14B-Q4_K_S.gguf, regardless, I tried both and still get the same error: KSampler The size of tensor a (19360) must match the size of tensor b (20240) at non-singleton dimension 1. Any ideas??? TIA, James