I built a multi-face + long-video accelerator for MiniMax H3 Face Refine
retained text
I’ve been using ComfyUI-H3-FaceRefine on MiniMax H3 videos, and I really like what it does for small faces. But I kept hitting the same annoying case: a shot has several people in it, and I want to clean up all of their faces. The normal approach is basically “pick a face, run it, repeat.” It works, but once the video gets longer, waiting through several full H3 passes starts to feel pretty painful. So I spent some time making a fork that packs multiple tracked faces into one atlas, runs them through H3 together, then puts them back where they belong. Here it is: [https://github.com/PullMyBoots/ComfyUI-H3-FaceRefine-Accelerated](https://github.com/PullMyBoots/ComfyUI-H3-FaceRefine-Accelerated) It can handle up to 9 faces in one pass, splits longer videos into smaller jobs, and carries a little latent context into the next chunk so the joins are less jarring. It’s not magic, though. If you put four faces into a 768 atlas, each one gets about a 384px cell. That’s a perfectly reasonable trade for a group shot, but I’d still use the original single-face path when I care about one person’s face as much as possible. On my 4090, a 15-second clip with four faces went from roughly 16.5 minutes of serial processing to about 3.5 minutes. That’s just one test on my setup, not a promise that everyone will see the same number. The old workflows should still load, and the accelerated nodes are optional. I mainly made this because I wanted to stop treating every group shot like four separate videos. Would love to hear whether this is useful to anyone else doing H3 video work — and especially whether people notice problems with identity or temporal consistency in real footage.
document fetched 2026-09-08 00:39 · canonical https://reddit.com/r/comfyui/comments/1w9yqxb/i_built_a_multiface_longvideo_accelerator_for