Open ComfyUI Workflow: Long Lipsync With MiniMax, Segmented to Avoid Degradation
martinerous · reddit · 2026-10-27
The author shares a ComfyUI workflow for long lipsync built on MiniMax H3: feed in speech audio, an optional reference image and variation keyframes, and it outputs a lipsynced talking-head video.
- Approach: a reference model (hybrid works too) plus custom nodes to seamlessly extend the video latent, randomly inject end frames for talking-head variation and to guard against degradation, and an audio splitter node to drive the process.
- Handles H3 latent-grid specifics (17k+5 → 5k+2, etc.) and includes a "retake" workflow that replaces the middle of an already-generated latent to fix broken moments.
- Looping: ComfyUI StartLoop becomes very slow to start with dozens of iterations (it unwinds the loop to build cache at once), so the author splits at fixed times with video-latent cross-fades and swaps the generated audio for the original track to avoid an audio VAE roundtrip.
- Splitting by silence doesn't work because it can't align with the H3 latent grid, causing choppy or misaligned audio.
- In theory it could inject different prompts at fixed timestamps, but complex long projects are better served by existing motion-context nodes.
More from coding & agent
- Anthropic engineer: build systems that prompt themselves — 90% of our engineers already do — Paimaamu · 2026-10-27
- golang/go repo flooded with a dozen AI-generated CVE security issues in minutes — evilsocket · 2026-10-27
- SuperPlane 1.0 launches open source: 52% of team's merged PRs were automated — Arindam_1729 · 2026-10-27
- OpenAI DevDay Berlin: Codex, Agents API and Terra 6.1 launch on site — paw_lean · 2026-10-27
- MiniMax M3.1 Flash hands-on: agent builds, tests and iterates a game end-to-end — Aiden_Tech_Ai · 2026-10-27
- Terminal color themes become a physical button that pings you when agents finish — dee_hw · 2026-10-27