Overview
Using Wan Animate 2.2, transfer motions from Reference Video to animate the character in Reference Image, or replace the character in Reference Video with the character in Reference Image. Tested with videos up to ~20 seconds long but theoretically should support unlimited length.
Still a WIP as there are some rough edges (certain reference videos and images work better than others, character identity drifts the longer a video goes), but releasing it as I don't see other similar workflows available on Civitai yet.
Key Features
Using
WanVideo Block Swap&WanVideo Animate Embeds, this workflow splits long videos into small "windows" of 81 frames (~5 seconds) so that theoretically unlimited video length can be supportedUsing RIFE VFI, this workflow interpolates the generated frames so that buttery smooth video of 60FPS or more (configurable in workflow) can be generated
Custom ComfyUI Nodes
Model Download Links
Important Notes
Generating at 480p (480 x 832 pixels), system RAM usage peaks at around 47.8GB and VRAM usage peaks at around 15GB so you would need a system with ≥16GB VRAM, ≥48GB RAM to run this workflow as-is.
You might be able to lower the system requirements by tweaking the various settings.
Description
Improved retention of facial details
Reduced RAM usage
FAQ
Comments (29)
Getting an error: When loading the graph, the following node types were not found:
WanVideoAnimateEmbeds . I've updated comfyui, wanvideowrapper custom node, but still getting this issue.
It worked fine with me, I can tell you which versions & stuff I have that might help you to get it working:
-((The most important I guess)) I have WANVideoWrapper (one of the very recent versions that came out this month),
-I don't anymore have MaTeZZ ComfyUI-WanVideoWrapper-MultiTalk , because it started to make Serious Conflicts with the latest WANVideoWrappers & KJNodes, btw those nodes latest versions already support Multitalk, so you can safely remove the MaTezz nodes.
-I use WAN Animate 14B KiJai Q8_0 (you can use any other quant or the safetensors version for sure)
-I use portable Comfy v.0.3.64 with frontend 1.29.3
** Make sure to have a recent WANVideoWrapper & KJNodes, I guess this should help you to get rid of the error with whatever Animate workflow you use later.
@hazzoom82659 I have updatedd comfyui and the wanvideowrapper from the manager, i'll try to make a clean install and clone the repo directly, see if i have better luck.
Where are we supposed to get the vitpose-l-wholebody.onnx and yolov10m.onnx models from?
I've updated the main post with links to model downloads
ComfyUI-WanVideoWrapper/wanvideo/modules/model.py", line 2057, in wananimate_face_embedding
self.motion_encoder.to(self.main_device)
^^^^^^^^^^^^^^^^^^^^^^
AttributeError: 'NoneType' object has no attribute 'to'
Help me !
I haven't encountered that error before, but here's what Perplexity suggests: https://www.perplexity.ai/search/comfyui-wanvideowrapper-error-2HyxZI9PR7Sj15c.Lg0uPg
Is this possible to use resolution more than 480p if i have 16GB VRAM?
Im struggling with workflows that allows to generate long videos, but only at low resolutions - when i go above 640x480, i'm getting VRAM allocation errors, regardless of requested frame count, fps and block swaps.
Official animate workflow from comfy templates, allows me do make videos in 1024x768 and even 1200x900 that are looking awesome, but they can have maximum 77 frames which is 4 seconds). Of course, they can handle more than 4 seocnds, but with terrible workaround - making batch of new separate videos, one by one, and connect them via first and last frame. It causes glitches and weird transitions that are not acceptable.
Is there any way that allows to make let's say 8 seconds video at 1280x720p?
depends how much ram you have, if you have enough ram the block swop should work, though i have 24gb of vram, im barely going over 16gb if i do while i rely on me 64gb of ram to help out. the hi9gher resolution defs make this workflow work better.
I want to make the character in the video undress. How should I get the ref_image?
Hmm I think you mean how to get the ref_video if you already have a photo of the character you want to undress? For that you'll probably need to find a video of someone undressing and use that as ref_video, then plug in the character you want to undress as ref_image
Awesome, thanks man! Out of curiosity, is there any chance you could share the ref_image from your demo? No worries if not!
@aaronchen233 Sure, I think this is the one: https://civitai.com/images/109890373
@wildkrauss Thank you so much.
Hello again.
Could you implement an IP ADAPTER in your workflow? That would be an interesting option. For example, at a resolution of 576x1024, certain discrepancies can still be seen in terms of the person's gaze. Don't get me wrong, your workflow is amazing. I also did a little test with the REACTOR extension. It's an improvement, but I've read on other forums that IP ADAPTER would be much more suitable in this regard. What do you think?
@drak0n that's actually something I've noticed too and been wanting to do as well, but as far as I can see there doesn't seem to be any ready-made solutions yet. Looks like there used to be a ComfyUI module called ComfyUI-IPAdapterWAN which works with Wan 2.1 (not 2.2), but the GitHub repo has been taken down and it's not available through the ComfyUI Manager either.
I'm thinking that for now the best options would be either:
1. Get a cleaner reference image through another workflow using IPAdapter, PuLID, HyperLoRA or something similar
2. Or straight-out train a Wan 2.2 LoRA of the character to animate
Neither of these options are straightforward, and #2 seems to be overkill unless you're planning to use the same character on multiple animations.
@wildkrauss I searched on Reddit and found something. I hope it helps. You can search Google for "PR-ComfyUI-IPAdapterWAN-1c65dfb4" or "reddit wan 2.2 video ip adapter." It's in the first few results. There are two possible solutions in that post.
@drak0n Thanks for the suggestion. Both of those repos seem to be "backup" copies of the previous ComfyUI-IPAdapterWAN which was taken down for some reason.
I've been trying to get it to work, but no luck so far. Here are my current findings:
1. The IPAdapterWAN node expects a model input loaded from either a UNET or Diffusion Model node, but this workflow uses a special WanVideo Model Loader in order to support block swapping and the model output is incompatible with the IPAdapterWAN model input
2. So I've tried plugging IPAdapterWAN into a simple I2V workflow, but it throws the error RuntimeError: mat1 and mat2 shapes cannot be multiplied (257x1280 and 1152x1280)". This suggests that the ip-adapter.bin model file used by the node is incompatible with Wan 2.2 in general. Both ChatGPT 5 and Claude Sonnet 4.5 confirm this finding.
Looks like we will need to wait for a Wan 2.2-compatible IP Adapter to be released.
@wildkrauss Thanks for your reply. Then we'll wait for something better. By the way, I notice that towards the end of the video, the colors change. The colors are not consistent throughout the video. I know there is a special node for colors called "Color Correct (mtb)" that could fix this. I hope I'm not bothering you with these discussions. Thanks in advance.
@drak0n No worries about the questions; keep them coming! I was away for a few days, but I've experimented with the Color Match KJ Node but got horrible results.
I'll try out the Color Correct (mtb) node and see if it works better.
@drak0n I've tried the Color Correct (mtb) node but it doesn't seem to help. But what I've noticed is that the color changing is very apparent when you use Wan2.1_I2V_14B_FusionX_LoRA; I haven't experienced noticeable color changes without it. Can you try removing it and see if the results are better?
@wildkrauss Indeed, that lora (regardless of the value you use) causes that color change.
It said [ONNXRuntimeError] : 1 : FAIL : Non-zero status code returned while running QuickGelu node. Name:'/model.0/act/Mul/QuickGeluFusion/' Status Message: CUDA error cudaErrorInvalidPtx:a PTX JIT compilation failed
is this a onnxruntime-gpu version issue?
which onnxruntime-gpu are you running on?
Could be. I'm running onnxruntime-gpu 1.23.2, Python 3.13.6, pytorch 2.9.0+cu130. Perhaps you can try upgrading onnxruntime-gpu?
Sorry for the stupid questions. In the workflow, there is a section called "Character Mask and Background Video Preprocessing." Does this section have an effect on the video, or does it need to be activated? Thank you in advance.
@drak0n that section is for the "Replace Character" mode, where the character in the ref_video is replaced with the character in the ref_image.
If you're only intending to use the ref_video as a driver to animate the character in ref_image, you don't need to bother with it.
I’m happy that you released your workflow. I tried generating a video using it, but no matter what I do, the face becomes distorted or the overall image quality turns out coarse.
The only settings I changed from the original workflow were the resolution (set to 544×720) and switching the model to Q8.
Why does the video end up in such low quality?
My input image is 1152×1536 and shows a woman from above the knees.
The reference video is a 544×720 dance clip of a woman, also framed from above the knees. (I also tested with a 1080p version, but the result was the same.)
Are there any other parameters I should consider? I would really appreciate any help from anyone who might know the cause.
@yakinamashake that's a weird one I haven't experienced, but I'm guessing that it might have to do with how Wan 2.2 internally processes videos.
From what I gather, Wan 2.2 has been trained on videos with 16:9 (or 9:16) aspect ratio and don't work so well with other aspect ratios. Can you try a 16:9 aspect ratio such as 1280x720 or 832x480 (the two officially recommended video sizes) and see if you get better results?
@wildkrauss Thank you so much for the suggestion. I switched to a 16:9 resolution as you recommended, and the results improved dramatically. Really appreciate your help!