沐阳 H3 · 五视图素材绑定(内部)
Bind a turnaround sheet to every shot instead of uploading it nine times
- media
- sheet
- 增强素材
- 增强分镜计划
Character consistency across a long generator chain is the single most complained-about problem in AI video, and the least glamorous fix is the one that actually works: give the model more than one view of the character. A single front-facing portrait gives H3 an identity to match and almost no information about what that identity looks like from behind. A turnaround sheet - front, side, back, three-quarter, and a face close-up - gives it geometry.
H3DirectorTurnaroundMedia takes a generated turnaround sheet and files it into the shared media inventory, then binds it to every shot in the plan. One image, referenced everywhere, instead of nine manual uploads you'll forget to update when the character's costume changes.
The mechanism
Four inputs:
media- theMINIMAX_H3_MEDIAinventory (Director or Media Agent).sheet- the turnaround IMAGE. The node keeps the first frame, files it in the catalog as a generated asset (generated_turnaround_sheet.png, described as a 角色五视图分镜参考 - a five-view storyboard reference), and continues its numbering from whatever's already in the bag.plan_json- the storyboard plan, as JSON. Not valid JSON means a clear error, not a crash somewhere downstream.owner_id(default empty) - the owning job/card identity, used so the binding knows which Director instance it belongs to.
Two outputs: 增强素材 (the inventory, now including the sheet) and 增强分镜计划 (the plan, now referencing it). Both go back into the Director.
The design point worth noticing: it appends to the shared inventory rather than attaching to individual shots. The shared area is the right home for anything used throughout a job - a character sheet, a location, a music bed - and the pack's media model lets each shot card use its own media or append the shared inventory. Bind at the shared level and every card inherits it.
What has to exist first
This node is glue, not a generator. The turnaround sheet itself comes from the five-view path, which lives in AIMixer's ComfyUI-MAINodes (H3ContactSheet, H3ContactSheetDecode, and friends) plus the five-view LoRA. Those are optional third-party packs that the Director only loads when the relevant switch is on - with the switch off, the core pack never imports them.
So the realistic setup order is: install MAINodes, get the five-view LoRA, generate a sheet you're happy with, then let this node bind it. If MAINodes isn't there, the Director's enhancement check fails before it starts sampling, which is the civilised way to find out.
Note also the plan's provenance: workflows store portable ComfyUI/input references and do not embed media payloads in the JSON. If you move a project to another machine, the sheet goes with you only if the file does.
Install
cd ComfyUI/custom_nodes
git clone https://github.com/civilcoco/ComfyUI-MiniMaxH3-Myang
Restart, hard-refresh, category 沐阳 H3/导演台/内部. The pack declares no Python dependencies; the media catalog is its own code. Alongside the H3 model, Qwen text encoder and both VAEs, you'll want a face detector and the MAINodes pack for the sheet generation that feeds this node.
Why bother
Because the alternative is drift, and drift is the thing that kills long chains. The pack says it plainly in its own limitations list: long chains accumulate losses in picture detail, timbre, brightness and saturation, and a clean seam does not guarantee that content quality stays constant down the chain. Anything that pins identity is worth the wiring, and a turnaround sheet is close to the cheapest identity anchor available - one image, referenced from every segment.
If you only take one thing from this node: a five-view sheet is not a nicety for character work, it's the difference between a character and a lookalike by segment 12. The character consistency and identity preservation write-ups go deeper on why reference count beats reference strength.
Inputs (4)
| Name | Type | Default | Description |
|---|---|---|---|
| media | MINIMAX_H3_MEDIA | — | |
| sheet | IMAGE | — | |
| plan_json | STRING | — | |
| owner_id | STRING | — |
Outputs (2)
| Name | Type | Description |
|---|---|---|
| 增强素材 | MINIMAX_H3_MEDIA | — |
| 增强分镜计划 | STRING | — |