WAN 2.2 VACE Fun! (Patreon)

Video - https://youtu.be/UR5phAU5tps

Overview

What's more fun that WAN 2.2 Fun? WAN 2.2 VACE fun! Why? Well, the VACE control layer has been created to give better character and style consistency, when compared to just plain old WAN Fun...

Being a Control layer you still get the capabilities of WAN fun, so control nets, first/last frame and the like, only this time it's as an extra layer on top of the WAN 2.2 Text-to-Image models.

From my testing, 4-steps seems to be just fine meaning with a modern GPU you're looking at under 1 min and 30 seconds for an 81-frame video, with the usual block swap and torch compile options available to help with VRAM usage and performance.

When providing reference images, it's a good idea to use images with white borders, like this example:

Robust Video Matting is one simple way to automatically remove the background from your video frames as you can change the bg_colour to white (instead of the usual green):

If you're providing a background image then the white borders aren't required, for example:

Performance options

Block swap balances VRAM/RAM. Alternative config:

Also, most loaders have their own balancers too, such as here with the T5 loader's "load_device" being set to "offload_device":

The prompt node can also be run on gpu or cpu, and offloaded as required, with a disk cache option available for speed.

Node Updates

Note that updated versions of Kijai's nodes now have an extra_model input in place of the vace_model. The VACE Module Select can be connected to the extra_model input in this case, as vace_model is no longer present.

See v5a for this example.

17/Sep/25 - This issue has now also been fixed with the updated nodes... control video no longer needs a start_image as well :)

Attachment Info

Both of the workflows used in the video are attached - the main Fun VACE on and the one I used to compare the various steps.

You'll likely just want to use the non-compare version, but it's been included just in case :)

The v5a workflow uses the newly updated model loader (which no longer has vace_model)

Model files can be downloaded from Huggingface:

https://huggingface.co/Kijai/WanVideo_comfy/

https://huggingface.co/Kijai/WanVideo_comfy_fp8_scaled

For custom nodes, the main ones are:

ComfyUI Wan Video Wrapper - https://github.com/kijai/ComfyUI-WanVideoWrapper

KJ Nodes - https://github.com/kijai/ComfyUI-KJNodes

Video Helper Suite - https://github.com/Kosinkadink/ComfyUI-VideoHelperSuite

RGthree - https://github.com/rgthree/rgthree-comfy

Controlnet_aux - https://github.com/Fannovel16/comfyui_controlnet_aux

Video Matting - https://github.com/Fannovel16/ComfyUI-Video-Matting