# MiniMax H3 Extender Working

## Summary

Chris's current working `MiniMax-H3` extension workflow for continuing an imported video with reference-driven follow-on scenes, review checkpoints, and recovery / re-assembly support.

## Owner

- `Chris`

## Status

- `working-custom`

## Source

- Personal workflow supplied by Chris via Discord attachment
- Imported into repo: `2026-08-15`

## What this workflow is

- Starts from an imported source video and audio pair
- Builds imported-video context for `H3` continuation rather than starting from a clean prompt-only clip
- Uses a plan-driven extension loop for multiple follow-on scenes
- Supports scene + character reference images, including a third character lane
- Includes a review gate and checkpoint save after each generated scene
- Includes muted-by-default recovery nodes for re-assembling a finished extension without re-rendering

## Key working settings captured from the graph

- Input video placeholder: `1.mp4`
- Input audio placeholder: `リアルもエーアイもこんなんばっか(ソウル).wav`
- Source timeline basis: `24 fps`
- Base scene resolution node: `1152x640`
- Main loop `H3 REF2VA` node resolution: `864x480`
- Clip length on the H3 generation nodes: `124` frames
- Steps constant: `8`
- `Boolean - Distilled`: `true`
- `Boolean 2 Pass`: `false`
- `HIGH QUALITY`: `false`
- Optional RTX VSR upscale branches are present but disabled behind switches
- Primary preview / loader override references `taeh3.safetensors`

## Reference structure in this graph

- `Picture 1`: environment / master scene
- `Picture 2-3`: Bill body + face
- `Picture 4-5`: Nancy body + face
- `Picture 6-7`: Becca body + face
- Imported source video and its audio are also fed back in as continuity references

## Notable workflow sections

- Imported media ingest via `LoadVideoUI` and `LoadAudioUI`
- Shared set/get lanes for `CLIP`, `VAE_Vid`, `VAE_Audio`, prompt, dimensions, and media assets
- Initial `MiniMaxH3ReferenceToVideo` block for reference-to-video generation from scene and character stills
- `MiniMaxH3ChainPlan` plus `MiniMaxH3ChainLoopStart` / `Current` / `Context` / `LoopEnd`
- `MiniMaxH3ChainReview` gate set to checkpointed review before continuing
- `MiniMaxH3ChainSegmentSave` for resume checkpoints
- `MiniMaxH3ChainManifestLoad` and alternate `MiniMaxH3ChainAssemble` for recovery / assemble-later flow

## Embedded project assumptions

- The prompt text is authored around named characters `Bill`, `Nancy`, and `Becca`
- The embedded example dialogue and action are tailored to a specific rooftop / shop continuation scene
- The graph expects local reference image filenames such as `Scn037.png`, `Char256_Turnaround.png`, `Char093_Headshots.png`, and related assets

## Why this lives under `custom/`

- This is not an untouched external import
- This is not yet a cleaned general-purpose master
- The graph contains Chris-specific prompt language, asset names, and workflow assumptions that are worth preserving exactly as a working baseline

## Likely required custom nodes

- `kijai/ComfyUI-KJNodes`
- `ethanfel/ComfyUI-MiniMaxH3-Contex-Loop`
- `WhatDreamsCost-ComfyUI`
- `DumasNodes.git`
- `comfyui-kjnodes`
- `comfyui_fearnworksnodes`

## Notes for future cleanup

- A future generalized version could move into `optimized/image-to-video/` after stripping Chris-specific prompts, filenames, and scene assumptions
- If this evolves further, keep the raw working graph here and promote only a cleaned derivative to the optimized layer