ComfyUI workflow for MiniMax H3 character swap: 64s 1080p video in 15 segments

wjc_5 · reddit · 2026-11-08

A detailed ComfyUI workflow for MiniMax H3 character swap / motion transfer, plus an agent skill that writes prompts and batch-runs segments — demoed by swapping characters across a 64-second clip cut into 15 segments.

Frame alignment: H3 needs 17n+5 frame lengths; the workflow pads with the last frame, then trims, outputting 24fps at exact input length.
Context continuation: for split long shots, load the previous segment's output — its last 22 frames guide the next segment's start for seamless stitching; shots cut at natural boundaries run independently.
Speed: Spark-H3 sparse attention + Spectrum forecast caching; LoRAs: lightx2v Ref2VA turbo v0.1 at 1.0 plus 600-step EMA turbo LoRA at 0.5, and akatz-ai's Character Swap LoRA. The author argues character swap is narrow enough that stacked accelerators cost little quality.
Key settings: 720p (0.9–1.0 MP) handles 15s per pass with TemporalWindow (6.5s per window); 1080p should be cut into 4–6s segments; disable spatial tiles in TemporalWindow; Spark-H3 requires RTX 30/40/50.
Prompts: single-shot clips work with a video-capable web LLM like Gemini plus a template; multi-cut clips use the skill — name segments (1, 2, 3 or 2-1, 2-2 for continuation), it detects hard cuts locally, writes per-segment prompts, runs resumably, then stitches with a side-by-side comparison and per-segment report. Works on RunningHub or local ComfyUI; Claude Code/Codex can use it via SKILL.md. At 2.0 MP each segment takes 6–10 minutes.

Original post →

ComfyUI workflow for MiniMax H3 character swap: 64s 1080p video in 15 segments

More from coding & agent