Anikin Labs MiniMax H3 Face Refinement v1.0 is a ComfyUI workflow focused on improving facial identity consistency in MiniMax H3 video generation.

The workflow generates the base H3 video first, then runs a separate Face Refinement stage that tracks the face across frames, processes the face region, and stitches the refined result back into the original video.

The main goal is to reduce identity drift, facial instability and visible changes in character appearance between frames.

What this workflow includes

• MiniMax H3 Reference-to-Video generation

• Dual Sampling structure

• Face tracking

• Per-frame face refinement

• Face identity reference

• Face stitch-back

• Base video output for comparison

• Final Face-Refined output

• Debug face-track preview

• Built-in model/download notes with installation paths

• Resolution / megapixel reference table

Recommended settings

The workflow has been tested primarily at 0.6 MP.

For 16:9:

• 0.4 MP → 864 × 480

• 0.5 MP → 960 × 544

• 0.6 MP → 1056 × 608

• 0.7 MP → 1152 × 640

• 0.8 MP → 1216 × 672

• 0.9 MP → 1280 × 736

• 0.98 MP → 1344 × 768

For 9:16, width and height are swapped.

Face Refinement preset

The current tested Face Refinement settings are tuned to preserve identity while reducing the “face mask” effect that can appear when the face is small in frame.

Recommended values:

• Small face denoise: 0.25

• Large face denoise: 0.35

• Face Track smoothing: 7

• Size smoothing: 15

• Mask dilation: 24

• Feather: 24

• Blend: 1.0

Important notes

Face refinement does not guarantee perfect identity preservation in every shot.

Best results are usually achieved when:

• the character's face is clearly visible

• the face is not extremely small in frame

• camera movement is not excessively fast

• the face identity reference is a clear portrait

• the face prompt focuses on preserving identity rather than beautifying or redesigning the face

Dynamic shots, strong head rotations, motion blur and very small faces may still produce some identity drift.

Workflow logic

OUTPUT A saves the original H3 generation before Face Refinement.

OUTPUT B saves the final Face-Refined result.

This makes it easy to compare the original generation and the refined version and see whether Face Refinement improves the shot.

About

Created and tested by Roman Anikin (Anikin Labs).

This workflow is intended as a practical identity-preservation pipeline for MiniMax H3 in ComfyUI and will continue to be improved in future versions.