forked from NikoDemon80/ComfyUI-H3-Motion-Context
-
Notifications
You must be signed in to change notification settings - Fork 27
Home
Ethanfel edited this page Aug 18, 2026
·
3 revisions
Create a multi-scene MiniMax H3 production with one reusable ComfyUI sampling body. Review every scene, retry mistakes, resume interrupted runs, and assemble the accepted result from disk.
This wiki describes the current version 0.5 workflow on
feature/0.5-workflow-ux. Saved 0.4 workflows and checkpoints remain supported.
- New installation: Getting started
- Not sure which JSON to open: Workflow chooser
- Something is not working: Troubleshooting
| I want to… | Go here |
|---|---|
| Make a multi-scene T2V, I2V, FL2V, or Ref2V video | Workflow chooser |
| Decide how one scene connects to the next | Continuity and audio |
| Use image, motion-video, or audio references | References and source media |
| Continue existing footage or use a source soundtrack | References and source media |
| Inpaint, outpaint, extend a clip, or bridge two clips | Masked editing |
| Retry a scene or recover an interrupted run | Review, resume, and recovery |
| Understand where a feature came from | Feature origins |
| Look up detailed Plan fields or architecture | Advanced reference |
- A complete scene Plan runs through one recursive H3 graph.
- Each scene becomes a recoverable checkpoint before you review it.
- Continuity is a clear incoming-scene choice: Cut, Guide, or an AV-prefix transition.
- Long source media stays file-backed and is decoded one scene at a time.
- Final assembly reads accepted scenes from disk instead of accumulating the entire production as an image tensor.
If this is your first run, use a maintained Normal workflow and keep its supplied settings. Studio, semantic anchors, Detail Guide, and Detail AV are optional or experimental paths—not prerequisites.