v3.0 Is Finally Here!
Entirely new layout with grouped step-by-step sections.
New detailed Notes section with links to essential model downloads.
Added Res4Lyf custom node pack for Sigma previews and access to more sampler/sheduler combinations.
Thanks to BillyWicked for their collaborative efforts in testing and refining this version!
The workflow is still fully capable of NSFW content if you wish to use it for those purposes. Feel free to share any of the videos you are able to generate using this workflow!
I HIGHLY RECOMMEND READING THE FULL INFORMATIONAL ARTICLE REGARDING THIS MAJOR UPDATE - https://civarchive.com/articles/21684/torstens-wan22-i2v-gguf-workflow-version-30
v2.0 - still simplified and very efficient
added more quality-enhancing features without impacting rendering speed!
v1.0 - basic workflow, extremely fast, not as many features
TL;DR
It's a great workflow, some have said it's the best they've ever seen. I didn't say that, but other people have. I think it's the best workflow anyone has ever uploaded. You know what we need on this platform? We need to Make Workflows Great Again!
Description
This is for Quantized versions of Wan2.2 i2v, you can use 480p or 720p in the workflow, whichever you prefer.
FAQ
Comments (37)
I will remind people yet again that they misunderstand VRAM and the model size. A video model should NEVER be installed in VRAM- indeed you want as little of it in VRAM as possible. To force this, disable smartmemory (which is anything but smart), tell ComfyUI you have low or no VRAM etc. Then place model clear nodes in key areas of a workflow if needed to prevent OOM or insanely slow render times. Your video model should be forced to stream from RAM across the PCIe bus as needed (and you really should have a minimum of 64GB of RAM).
My video card has 16GB of VRAM, and I can use ANY model size (no need for silly quality killing tiny quants) because I force the model to stream. Does this have a performance impact? MY ancient PCIe bus does 16GB/sec- do the maths. Actually ComfyUI is so badly written, there is a small advantage when more of the model is in VRAM, but this might improve a 22s/it to 21s/it- ie., nothing that matters.
The only thing that should exhaust your VRAM is the actual working data set for the video frames- so at some point your output rez is too large, or you have too many frames for your VRAM amount- and there's no fixing that problem (render times will become exponentially slower when this happens).
Do you have a workflow that works this way? I'd love to try it.
Big words blobby, workflow or it didn't happen.
To be fair, I sort of see your point…however, many AI content creators have seen that the quantized versions don’t take away a huge amount of quality. It’s great that your specific PCiE can do 16GB/s but what I’m doing here is helping people with limited resources and experience to at least begin their journey of AI content creation.
Blobby, please share your workflow file
我的显卡只有10G显存，能使用吗
是的，你可以使用低容量显卡。我推荐Q3或Q4型号。(Hopefully Google Translate is correct on this)
Impressive work, I have the same GPU as you! I'm however not as smart as you though, it really amazes me how people whip up workflows. Thanks and I wish you all the best.
Thank you! This comment made my entire week! I hope you enjoy using my workflow and I will be updating it with a new V1.1 soon based on further tests (probably this weekend, or early next week).
100%. Blows my mind how they know how to configure it so well.
I'm new to this, always got blurry static, essentially nothing out of i2v workflows in the past week but this just worked with 4 loras on the first try.
I've got this error after first try: KSamplerAdvanced
Given groups=1, weight of size [5120, 36, 1, 2, 2], expected input[1, 32, 25, 90, 90] to have 36 channels, but got 32 channels instead.
Any reason for that? i didn't change anything in the workflow, just uploaded my image and left only light lora without bypass
Edit:
In case someone is having similar problem i solved it by unistalling following custom node flow2-wan-video. Then deleted the folder completely from custom-nodes folder just in case. After that WAN2.2 started to work.
Are you using Wan2.1 VAE Safetensors? These 14B Wan2.2 models still need 2.1 VAEs. Also make sure that you select "Wan" from any dropdown that gives you a model option, usually those are found in the nodes of the first step (first column of this WF)
In case someone is having similar problem i solved it by unistalling following custom node flow2-wan-video. Then deleted the folder completely from custom-nodes folder just in case. After that WAN2.2 started to work.
Nerdula oh I heard about that error. I never encountered it myself (most likely due to running all of this on a clean install of the ComfyUI Win Portable) - thank you for letting me know so I can advise more folks down the line if they have the same issue.
Still getting this error though. Don't have flow2-wan-video installed
Erosire This error is typically caused by three things:
1 - Using the Wan2.2 vae instead of the wan2.1 vae on the i2v models
2- Not having comfyui up-to-date (click Manager and ensure the version at the top of the manager screen >= V3.35)
3 - Using the flow2-wan-video module apparently (not sure, I don't use it but it seems like it caused the issue for some people)
Great workflow, thanks. After some testing I can offer a slight improvement. Change the sampler scheduler combination to LCM/Simple which I believe is the recommended configuration. I was getting some distorted faces during movement using Euler/normal but I regenerated the same scene (same seed) using LCM/Simple and the distortion was gone :)
I've also switched to UniPC/Normal due to the Euler issues. I heard that Euler was improved with Wan2.2, which is accurate. However, other samplers are still superior as you mentioned.
Thanks for the workflow!
Thank you for the Buzz!!
+1 -- I had a hard time figuring out ComfyUI for the first time, but you got me started!
bolthorseman glad to hear!
I don't understand how i can get good results. The workflow is working but my outputs have "generation artifacts" or more specifically "pixelation artifacts", it's glitchy, distorted, and blocky visuals. I'm just starting in all this and I don't understand what the problem
I'll have to see what selections you have in each section to understand how to determine what's wrong. Can you provide more information?
Still getting this error
Given groups=1, weight of size [5120, 36, 1, 2, 2], expected input[1, 32, 25, 90, 90] to have 36 channels, but got 32 channels instead
nvm, got it working. Need up to update ComfyUI to latest version.
Experimenting with the workflow it's just generating a ton of noise. At first it's the image, and then all frames after are just a ton of blurry noise.
I disabled all the loras, tried different samplers, changed cfg to 3.5. Still getting a lot of noise and blur. I'm using the 14B_Q4_K_M.gguf hi & low models. 2.1 VAE.
make sure you have the correct selections on both passes (high noise and low noise) for the sampler and schedulers, as well as the noise generation options.
High Noise should have the top row noise generation enabled, and the bottom row leftover noise return enabled also.
Low noise should have both of those noise options set to Disabled.
Also, both passes need to have the same scheduler and sampler selected.
TorstenTheNord Heya, I added some more steps and it started generating a better video but still a lot of noise. I tried another workflow using the fp8 scaled_kj.safetensor models and 20 steps and it worked fine. Just takes a lot longer than your workflow.
In your workflow my sampler settings are:
High : add noise=enable, steps=12, cfg=1.0, euler, normal, steps 0-3, return=enable
Low : add noise=disable, steps=12, cfg=1.0, euler, normal, 3-10k, return=disable
Clip is the e4m3fn_scaled. Wan2.1 VAE.
Is it supposed to be bullerwin's models? https://huggingface.co/bullerwins/Wan2.2-I2V-A14B-GGUF/tree/main
bamamahama First I would try with 12 steps to put the Step Start at 0 and Step End at 6 for High-Noise, then start at Step 6 for Low-Noise. I also find that LCM/simple or UniPC/normal are better sampler/scheduler combinations to prevent artifacts.
I have personally been using Bullerwin's models for this workflow.
Also, if you're using LightX2V, Causvid or Accvid LoRAs, you'll be better off keeping it to 6 steps at CFG=1. If you add more steps, make sure it's an even number, so the High Noise can take the first half of the steps, and the Low-Noise takes the last half. 50-50 split is very important here, in my experience with testing.
I get the error about 32channles. i dont have flow2 nodes, i got the latest version of comfy and using wan2.1 vae.
how to fix? Wan2.2 works with kijai nodes.
In your ComfyUI folder, depending if you have the Windows Portable one (which I highly recommend to prevent dependency issues) make sure to click the option "Update Comfy AND DEPENDENCIES" after you've updated the ComfyUI itself. If that doesn't work, you'll probably need to completely delete and reinstall ComfyUI (using the Windows Portable installation method).
I assume you're on Windows 10 or 11, so if you go with the the Delete & Clean Install method, do the following in this order:
1. Make a separate folder as a backup of your most important stuff like reference images/inputs, generated outputs, diffusion models, and LoRAs - but delete all custom nodes and everything else that could be causing dependency issues.
2. Use the RUN command from the Windows start menu (just click Start, type "run", hit Enter), then run %appdata% and delete the ComfyUI folder from the AppData system.
3. Perform a clean install ComfyUI, there are multiple ways to do it, found on the following link, just scroll down until you see "Windows Portable" - https://github.com/comfyanonymous/ComfyUI
this has been the best workflow for me so thank you so much! , very excited for the new version? how soon you think it will be?
Just added V2.0 information to the model page! Go ahead and try it (read the full article linked in the updated description) and let me know what you think!
TorstenTheNord Hi thanks wow. Excited to try, can you please link me the lightning lora you used for this one?
TorstenTheNord nvm got it from article cheers man.