Faces far from the camera turn to mush in MiniMax H3 videos. The open-source ComfyUI plugin ComfyUI-H3-FaceRefine fixes that. It crops each small face out, enlarges it, has H3 redraw it, then pastes it back into the original clip. It's MIT licensed, has over 600 stars on GitHub, and the latest version is 1.1.2.
Why Faces in the Distance Break
The plugin's author explains it clearly. H3 draws a face badly based on how much of the frame it fills, not on the output resolution. So even at 720p or higher, a face that's only a few dozen pixels tall still falls apart. Upscaling the video to 2K won't save it, because you're enlarging a face that's already been drawn wrong.
How the Fix Works
The process has four steps, all running automatically inside ComfyUI:
- Find and track the face. It detects faces on every frame and locks onto the same person, so the box follows them without jumping around.
- Crop to a square. It cuts out the face and enlarges it so the face fills a canvas of about 768×768.
- Have H3 redraw it. It regenerates that square at a low redraw strength. The smaller the face, the more it gets redrawn. A face that's already clear only gets a light touch.
- Paste only the face back. It puts the redrawn face back into the original video and leaves everything outside the face untouched.

The image comes from the project's GitHub repo, which has a before-and-after example. The left side is the original clip and the right side is the fixed one.
The idea borrows from FaceDetailer, a long-standing ComfyUI node that fixes faces in still images. This version works on a whole video instead of a single picture. One user who tested it on an RTX 4080 with a redraw strength of 0.6 reported that a 6.6-second clip took about two minutes to fix.
What Else It Can Do
- Fix only one person when there are several in the frame. Give it a reference photo of that person and it recognizes them by their face, so it stays on track even in a crowd. You can also pick the person by hand, shot by shot, in the interface.
- Handle videos with cuts. When scene-cut detection is on, it splits the video into shots and picks the person again for each one. You can also set it to skip shots where that person doesn't appear, which saves time spent on shots it doesn't need.
- Lip sync. H3 generates sound and picture together. If you connect the audio node the author recommends, the mouth still follows the voice when the face is redrawn.
- Anime characters. The default face detector is built for real people. For anime, switch to an anime face detection model, and use a recognition model made for anime as well.
What You Need
| Requirement | Notes |
|---|---|
| ComfyUI | Use a version recent enough to include the MiniMax H3 node (H3 is now built into ComfyUI) |
| H3 models | If you already make H3 videos, you have these |
| Face detection model | For example face_yolov8m.pt. This is the only file you have to add yourself |
| Three companion plugins | Impact Pack, VideoHelperSuite, and H3 NativeAudioLock. The example workflow needs them |
| VRAM | The example includes a compressed model set. With it turned on, H3 runs on 12 GB of VRAM |
After installing, you'll find two ready-made templates under Workflow → Browse Templates in ComfyUI. One picks faces automatically by a rule (the largest face by default). The other lets you choose the face yourself.
Have Your AI Agent Install It
I want to use the open-source ComfyUI-H3-FaceRefine to fix faces far away in my MiniMax H3 video: https://github.com/Carasibana/ComfyUI-H3-FaceRefine
1. Read its README first. Check whether my ComfyUI version includes the MiniMax H3 node and how much VRAM my GPU has. Tell me if it will run and whether I should use the compressed model.
2. List the plugins to install and the models to download, and the folder each one goes in. Wait for my OK before installing anything.
3. After installing, open its bundled template. Try a short clip of about 5 seconds from me first, using the README's default settings.
4. Capture the same frame from the original clip and the fixed clip, and show me the comparison. Tell me how long that run took.
5. Don't change my original workflow or model files.
Things to Know Before You Start
- Longer videos take longer. Run time scales roughly with the canvas area times the number of frames. If the automatic canvas is too big for your VRAM on a long video, it doesn't throw an error. It quietly switches to system memory, which can be ten times slower. Test with a short clip first.
- Frame counts get padded. H3 only accepts frame counts from a fixed set: 5, 22, 39, and so on. Other lengths get padded up to the next value. The extra frames are fixed and then dropped, so it only costs a bit more time.
- Reconnect old workflows. Version 1.1.0 changed how the redraw strength node is wired. If you open a workflow saved in 1.0.0, it will report a missing input. Reconnect the model line as the README describes.
- It reads videos from any path on your machine. If you don't want it touching files outside the ComfyUI folder, set the environment variable
H3_FACEREFINE_STRICT_PATHS=1. - It fixes faces only, not hands or bodies. The plugin can't fix warped fingers in the distance or distorted clothing.
References
- ComfyUI-H3-FaceRefine (GitHub)
- MiniMax H3 ComfyUI model files (Hugging Face)
- If you want to upscale a whole H3 video to HD, see Have your AI agent upscale AI video to HD


Comments (0)
No comments yet. Start the conversation.