Nodes/ComfyUI_MieNodes/Load Scene Asset ๐Ÿ‘
ComfyUI Node

Load Scene Asset ๐Ÿ‘

Four Outputs for Keeping the Room the Same Room

By MieMieeeeeยทCreated 2 years agoยทUpdated 2 days agoยท 269
Load Scene Asset ๐Ÿ‘
    • name
    • description
    • image
    • asset
    โ—„library_dirโ–บ
    โ—„name๏ผˆไธ‹ๆ‹‰้€‰ๆ‹ฉ่ต„ไบง๏ผ‰โ–บ

    Search the node menu for Load Scene Asset ๐Ÿ‘ and you'll come up empty: it's a hidden compatibility node, retired in the merge that collapsed the per-kind asset nodes into one Load Asset ๐Ÿ‘ with a kind dropdown. Old workflows still load it and it still works. New graphs should use the unified node with kind = scene.

    It's worth knowing what it does anyway, because scenes are the half of the asset library people under-use.

    A scene is a place, not a person

    The library's scene model is deliberately minimal: a name, one authored English description, and one picture. That's it - no audio, no voice line, no multi-image sheet. The description covers space, furniture and what stands where, lighting, atmosphere. On disk it's scenes/<name>/scene.json plus images/01.png.

    Four outputs come back: name and description as strings, image as an IMAGE, and asset as a MIE_ASSET bundle. Note what's absent compared to a character load - no voice, no audio. A scene has nothing to say.

    The asset wire is the practical one. Fed into a scene_N socket on MiniMax H3 Loop Plan Generator V2, the scene's authored description goes into the plan as an environment anchor (the manifest marks it environment, not identity) and the picture joins the reference batch as one of the <Picture N> entries - without consuming a <Subject N> slot, which is the part that matters if you're already tight against the nine-picture cap with characters.

    That division is the reason to bother. Character slots describe who, scene slots describe where, and keeping them separate is what stops the model from hallucinating a new kitchen every time the camera moves. Consistency across shots has never really been a prompt-wording problem - it's a matter of what the model can see while it generates.

    Install

    cd ComfyUI/custom_nodes
    git clone https://github.com/MieMieeeee/ComfyUI-MieNodes
    pip install -r ComfyUI-MieNodes/requirements.txt
    

    ComfyUI Manager works too: search ComfyUI_MieNodes. No weights, no model downloads; the library is a folder of PNGs and JSON. Restart after installing so the hidden node registers, and note that library_dir in the input list is ignored at execution - the root is fixed per machine (the plugin's asset_library/, or $MIE_ASSET_LIBRARY), reachable through the ๐Ÿ“‚ button.

    Things that go wrong

    An empty name dropdown means no scenes in the library - remember the dropdown lists scenes only, so a library full of characters looks empty here. Use Save Scene Asset (or Save Asset with kind = scene) to create one.

    does not fit the simple model (one image) fires when the folder holds more than one PNG. Trim to one picture.

    If a scene loads fine but the video still invents a different room, the picture is doing its job and the description probably isn't: state spatial facts - what stands where, where the light comes from - rather than mood words. Vague atmosphere text gives the model permission to improvise.

    Categoryhidden

    Inputs (2)

    NameTypeDefaultDescription
    library_dirSTRINGโ€”
    nameCOMBO๏ผˆไธ‹ๆ‹‰้€‰ๆ‹ฉ่ต„ไบง๏ผ‰1 options: ๏ผˆไธ‹ๆ‹‰้€‰ๆ‹ฉ่ต„ไบง๏ผ‰

    Outputs (4)

    NameTypeDescription
    nameSTRINGโ€”
    descriptionSTRINGโ€”
    imageIMAGEโ€”
    assetMIE_ASSETโ€”