Put yourself into any dance video – swap up to 5 people at once, 100% local (MiniMax H3 + ComfyUI)
retained text
Left: my source video. Right: three of the five dancers swapped for my own characters; the other two are untouched. H3 gets confused when people cross or overlap: it puts the new face on the wrong person, or one person ends up in another's clothes. My fix in H3MD 0.5 is to find those moments first and give the model a keyframe for each one. **How it works** - **Keyframe workflow (Qwen Image 2.1 edit, 2 steps).** Drag in the source video, add a picture per character, write "picture 1 → left 2" and one description line per person. It tracks the people (SAM3.1), picks the moments where they cross or overlap, redraws just those people and packs everything into a plan. About 2.5 min for a 26 s clip. - **Then the Director renders the swap** with the Remake LoRA, the source's own sound locked. 26 s took 29 min on an RTX 5090. - **Lesson learned:** without a prompt enhancer, the edit model follows your text. One simple line per person is enough, but if the line contradicts the picture (wrong gender, ethnicity or clothes), you get the wrong person. - **Remake only works at 1664×960 / 960×1664** (or 2560×1472). 0.5.5 switches to it automatically. **Also in 0.5: face refine for wide shots.** Small faces (50–100 px) are redrawn with H3 inside a SAM3 head mask, then VOSR2. Much clearer, not perfectly sharp; keep faces out of very wide shots if you can. Workflows (free): https://civitai.com/models/3037225?modelVersionId=3478241 Nodes: https://github.com/1113800942/comfy-h3md-director Walkthrough (Chinese voice, English subtitles): https://youtu.be/WMQJjmHZikc
document fetched 2026-11-19 00:42 · canonical https://reddit.com/r/comfyui/comments/1xzrra6/put_yourself_into_any_dance_video_swap_up_to_5