Nodes/ComfyUI-H3-Cast/H3 Scene Asset (Save/Load)
ComfyUI Node

H3 Scene Asset (Save/Load)

Save a location once, prompt it by name in MiniMax H3

By kat3ri·Created 21 days ago·Updated about 23 hours ago· 2
H3 Scene Asset (Save/Load)
  • image_1
  • image_2
  • image_3
  • image_4
  • scene_asset
  • views
  • report
operationLoad
saved_scene
name
description
confirm_updatefalse

MiniMax H3 will hold a room, a street, or a set dressing consistently across shots - but only if you actually feed it the reference images. That's the whole job of this node. H3 Scene Asset (Save/Load) is the "places" half of the H3-Cast pack: save a location once with up to four view images and a description, then pull it into any future workflow by picking its name from a dropdown. No re-connecting raw image references, no hunting through old graphs to remember which render was which.

Think of it as the sibling of the pack's H3 Cast Member node - same idea, but for sets instead of actors. One difference: a scene gets no voice clip (a location doesn't talk), and its description is written for setting, mood, and key details rather than appearance.

How it works

Like everything in this pack, it adds no new conditioning mechanism. It saves assets to disk and hands them to the stock MiniMaxH3ReferenceToVideo node, so every native H3 knob keeps working.

Each saved location is a plain folder - <comfy user dir>/default/minimax_h3/scene_sheets/<name>--<id>/ - holding a manifest.json plus images/*.png. Nothing is embedded in the workflow file, so assets survive restarts, reinstallations, and graph edits. The on-disk format is deliberately compatible with ethanfel's ComfyUI-MiniMax-H3-Guide reference sheets, so a location saved by either pack loads in both. You can point the whole root elsewhere with the MINIMAX_H3_SCENE_SHEETS_DIR environment variable if you want your library on a separate drive.

When you later Load it and plug the scene_asset output into H3CastToVideo, the compose node fans your saved views into H3's nine reference-image slots, writes the <Picture 1>, <Picture 2>: ... intro line with correct tag numbering, and your prompt just says "Maya sits by the window" - location and character references are handled for you.

The inputs that matter

Three modes, one dropdown (operation): Create, Load, Update. The rest of what you'll actually touch:

  • name - required in Create mode; becomes the folder slug and the dropdown label.
  • description - used verbatim in the auto-written prompt intro. This is your setting/mood text, so write it like you'd write the start of a prompt, not like a file-name caption.
  • image_1 through image_4 - your view images. One is enough; up to four.
  • confirm_update - a safety latch. In Update mode nothing gets appended unless this is on.

Forget a name or an image in Create mode and the node raises a clear error rather than writing a half-empty folder. Note that saved_scene is unused in Create mode - the "(none -- unused in Create mode)" option always sits at the top of that dropdown on purpose, so Create graphs stay valid when shared between machines with different saved sets.

Outputs

  • scene_asset - the bundle; wire this into H3CastToVideo's scene_asset input (or the Extend variant).
  • views - a preview batch of all saved views, resized to the first view's canvas so they tile in one preview image.
  • report - a short string saying what happened, e.g. Created 'cafe' (3 views) or Updated 'cafe' (+1 views).

Install

ComfyUI Manager works - search "ComfyUI-H3-Cast" - or do it by hand:

cd ComfyUI/custom_nodes
git clone https://github.com/kat3ri/ComfyUI-H3-Cast

Then restart ComfyUI; nodes appear under MiniMax H3/cast. The pack needs a ComfyUI build with the native MiniMax H3 nodes (comfy_extras/nodes_minimax_h3.py), and that's about it - no extra Python dependencies, no model downloads of its own.

Where people get burned

  • Update does nothing until you flip confirm_update. That's deliberate, not a bug - it's the write-protect so a stray graph run doesn't mutate your library.
  • Nine views is the ceiling, because H3 has nine image-reference slots shared across everyone and everything in the shot. Try to append past it and the node refuses.
  • Keep view aspect ratios consistent. The preview batch is squeezed to the first view's canvas, so a mix of 16:9 and square stills looks wrong in the preview - and mismatched references also confuse the model.
  • The real gate is the license, not the install. MiniMax H3's community license excludes the US, EU, UK, and South Korea from running the local weights at all. This node is a thin layer on top of those weights; if you're in an excluded territory, none of this is legal to run, and no setting fixes that.
CategoryMiniMax H3/cast

Inputs (9)

NameTypeDefaultDescription
operationCOMBOLoad3 options: Load, Create, Update
saved_sceneCOMBO1 options: (none -- unused in Create mode)
nameSTRING
descriptionSTRINGLocation description used verbatim in the auto-written prompt intro -- setting, mood, key details.
image_1optIMAGE
image_2optIMAGE
image_3optIMAGE
image_4optIMAGE
confirm_updateoptBOOLEANfalseSafety latch: Update appends connected images to the saved location only when this is on

Outputs (3)

NameTypeDescription
scene_assetH3_SCENE_ASSET
viewsIMAGE
reportSTRING