Simple Self-Forcing Wan1.3B+Vace workflow

This is a very simple workflow to run Self-Forcing Wan 1.3B + Vace, it only uses a single custom node, which everyone making videos should have: Kosinkadink/ComfyUI-VideoHelperSuite. Everything else is pure comfy core.

You'll need to download the model of your choice from here lym00/Wan2.1-T2V-1.3B-Self-Forcing-VACE · Hugging Face, and put it inside your /path/to/models/diffusion_models folder.

This workflow can be used as a very good start for experimenting. You can refer to this [2503.07598] VACE: All-in-One Video Creation and Editing for how to use Vace. You don't need to read the paper of course, the information you are interested in is mostly at the top of page 7, which I reproduce in the following:

Basically, in the WanVaceToVideo node, you have 3 optional inputs: control_video, control_masks, and reference_image.

control_video and control_masks are a little bit misleading. You don't have to provide a full video. You can in fact provide a variety of things to obtain various effects. For example:

• if you provide a single image, it's basically more or less equivalent to image2video.

• if you provide a sequence of images separated by empty images: img1, black, black, black, img2, black, black, black, img3, etc... then it's equivalent to interpolating all these img, filling the blacks. A special case of this one to make it clear is if you have img1, black, black, ..., black, img2, then it's equivalent to start_img, end_img to video.

• control_masks control where Wan should paint. Basically if wherever the mask is 1, the original image will be kept. So you can for example pad and/or mask an input image, like this:

  and use that image and mask as control_video and control_mask, and you'll basically do a image2video inpaint and outpaint.

• If you input a video in control_video, then you can control where the changes should happen in the same way, using control_mask. You'll need to set one mask per frame in the video.

• if you input an image preprocessed with openpose or a depthmap, you can finely control the movement in the video output.

reference_image node is basically an image that you feed to Wan+Vace that serves as a reference point. For example, if you put the image of someone's face here, there's a good chance you'll get a video with that person's face.