ComfyUI Wan VACE Video Extender

Github

This workflow uses Wan VACE to generate new frames from a specified point in a video.

What it Does

This is a lightweight, (almost) no custom nodes ComfyUI workflow meant to quickly extend a video with Wan VACE and a minimum of fuss. Just load your video, specify the extend start point and the number of new frames to generate.

Dependencies

My ComfyUI-Wan-VACE-Prep custom node package is required for this workflow. It replaces a large amount of awkward spaghetti workflow math, making this lightweight workflow version possible.

In ComfyUI Manager, search for Wan VACE Prep or just load the workflow and then visit the Missing tab in Manager.

This is a very lightweight package with no dependencies, so it is highly unlikely to break your system, if that is something you worry about.

I have not tested this workflow or the custom nodes under the Nodes 2.0 UI.

Configuration and Models

You'll need some combination of these models to run the workflow. As already mentioned, this workflow will not run properly on your system until you configure it properly. You probably already have a Wan video generation workflow that runs well on your system. You need to configure this workflow similarly to your generation workflow. The Sampler subgraph contains KSampler nodes and model loading nodes. Have your way with these until it feels right to you. Enable the sageattention and torch compile nodes if you know your system supports them. Just make sure all the subgraph inputs and outputs are correctly getting and setting data, and crucially, that the diffusion model you load is one of Wan2.2 Fun VACE or Wan2.1 VACE. GGUFs work fine, but non-VACE models do not.

• Wan 2.2 Fun VACE

  • bf16 and fp8

  • GGUF

• Wan 2.1 VACE

  • fp16

  • GGUF

• Kijai’s extracted Fun Vace 2.2 modules, for loading along with standard T2V models. Native use examples here.

  • bf16

  • GGUF

Troubleshooting

• The size of tensor a must match the size of tensor b at non-singleton dimension 1 - Check that both dimensions of your input videos are divisible by 16 and change this if they're not. Fun fact: 1080 is not divisible by 16!

• Brightness/color shift - VACE can sometimes affect the brightness or saturation of the clips it generates. I don't know how to avoid this tendency, I think it's baked into the model, unfortunately. Disabling lightx2v speed loras can help, as can making sure you use the exact same lora(s) and strength in this workflow that you used when generating your clips. Some people have reported success using a color match node before output of the clips in this workflow. I think specific solutions vary by case, though. The most consistent mitigation I have found is to interpolate framerate up to 30 or 60 fps after using this workflow. The interpolation decreases how perceptible the color shift is. The shift is still there, but it's spread out over 60 frames instead over 16, so it doesn't look like a sudden change to our eyes any more.

• Regarding Framerate - The Wan models are trained at 16 fps, so if your input videos are at some higher rate, you may get sub-optimal results. At the very least, you'll need to increase the number of context and replace frames by whatever factor your framerate is greater than 16 fps in order to achieve the same effect with VACE. I suggest forcing your inputs down to 16 fps for processing with this workflow, then re-interpolating back up to your desired framerate.

• If you can't make the workflow work, update ComfyUI and try again. If you're not willing to update ComfyUI, I can't help you. We have to be working from the same starting point.

• Feel free to open an issue on github. This is the most direct way to engage me. If you want a head start, paste your complete console log from a failed run into your issue.

Changelog

• v1.0.0 Initial release.