Redditor open-sources a simple workflow for infinite lipsynced talking head videos with MiniMax H3
martinerous · reddit · 2026-10-27
Reddit user martinerous released a ComfyUI workflow for generating arbitrarily long lipsynced talking head videos with MiniMax H3: feed in speech audio, an optional reference, and variation keyframes, and it outputs a lipsynced video.
Key techniques:
- Seamlessly extends the video latent via custom nodes, with random endframe injection for expression variation and degradation protection;
- Custom math to align with H3's peculiar latent grid (the 17k+5 → 5k+2 constraint);
- Fixed-time audio splitting instead of silence-based (silences can't align to the latent grid), plus latent cross-fade and swapping back the original audio to skip audio VAE roundtrip losses;
- A companion "retake" workflow to replace the middle of an already-generated latent for fixing glitched segments;
- Notes that ComfyUI StartLoop becomes very slow at 100 iterations because it unwinds all loop caches at once, and includes a looping workflow for when that's fixed.
The demo video was one-shotted; the author notes the approach could extend to timed prompt injection, though motion-context all-in-one nodes may suit complex long projects better.
More from coding & agent
- The 17-skill AI engineer checklist for 2026 — obsolete half by the time you finish — JafarNajafov · 2026-10-27
- The AI engineer job in 2023 vs 2026: one viral blank line — JafarNajafov · 2026-10-27
- DeepSeek Flash impresses dev: let it build its own tools, cheap enough to experiment — victormustar · 2026-10-27
- Solo dev builds a remote MCP server that runs a small business's website and socials — Ok_Cucumber_131 · 2026-10-27
- Agent loop on OpenClaw stays stable for over a week straight — vincent_koc · 2026-10-27
- Teaching Grok Bot to take your work: five ways to hand off tasks, from drafting to cross-app workflows — Scobleizer · 2026-10-27