ComfyUI fork cuts MiniMax H3 group-shot face refinement from 16.5 to 3.5 minutes
This_Temporary_8537 · reddit · 2026-09-08
A developer forked H3 FaceRefine for MiniMax H3 video generation to fix its biggest pain point: group shots required one full face-refinement run per person. ComfyUI-H3-FaceRefine-Accelerated tracks multiple faces, packs them into one shared atlas, runs H3 once, and pastes results back.
Key points:
- Long clips are split into smaller jobs with a small amount of latent context passed forward, avoiding visible chunk seams
- Trade-off: lower per-face resolution — four faces in a 768 atlas means 384px cells; close-ups or identity-critical faces should still use the original one-at-a-time workflow
- Benchmarked on a 4090: a 15-second, four-person clip took 16.5 minutes serial vs 3.5 minutes accelerated (4.7x)
- Repo includes demo clips, workflows, and benchmark details
Related event: Open-Source Fork Speeds Up MiniMax H3 Face Refinement by Nearly 5x(2 posts)→
More from coding & agent
- LangChain ships Managed Deep Agents 0.9, walkthrough from Interrupt London keynote — LangChain · 2026-10-26
- Knowledge Worker Kit adds a full section on organizing projects — dSebastien · 2026-10-26
- Spring AI 2.1.0-RC1 adds tool-call confirmation, pausing agent loops before real-data access — therealdanvega · 2026-10-26
- METR red-teams OpenAI's agent monitoring, finds bypasses and gaps — thione · 2026-10-26
- OpenAI: coding agents weakened safety monitoring, one lied about it — thione · 2026-10-26
- ConvFill: A 0.6B Talker Buys Time While a Large Reasoner Whispers the Real Answer — andimarafioti · 2026-10-26