One continuous 39-second shot out of MiniMax H3 segments of 4-6 seconds, made on an RTX 4070 SUPER 12 GB. Face, outfit and colour hold to the end. Most seams don't show, but one still does: at 0:30.7 her hands jump from the table to the cup in a single frame, in both videos. The two videos here are the same cafe scene, photoreal and comic, 8 segments each.
How it works: each segment continues from the latent tail of the previous one (FeiHou Easy H3, 接续 continuation, 22+22 frames), so motion carries over the seam. Segments 2 to N run in reference mode with the face reference and the establishing image, and every segment text repeats the full outfit and describes the props by their state.
In the archive:
- UI workflows for segment 1 and segment N, with a note inside each on what to set;
- the same graphs in API format;
- the planner system prompt (paste it into any chat assistant and get the segment texts) and the plan of the cafe scene;
- scripts: plan linter, chain runner over the ComfyUI API, seam and freeze checks, face similarity, review sheets.
Needs: ComfyUI with FeiHou Easy H3 and FeiHou Toolbox, KJNodes, MiniMax H3 Remix v0.8, the H3 text encoder and VAEs from Comfy-Org/MiniMax-H3, the H3 latent upscaler. ReActor only if you swap the face on the first frame.
Time on a 4070 SUPER with 96 GB RAM: about 1.5 minutes for segment 1, 2.5-3.5 minutes for every continued segment.
The full method, by hand or with an agent: https://civarchive.com/articles/38141
How a local agent runs this workflow by itself on the same 12 GB card, with its own brain switched off while ComfyUI works: https://civarchive.com/articles/38178
Description
First release: the cafe scene plan, UI and API graphs for segment 1 and segment N, planner prompt, scripts.

