Nodes/VRGameDevGirl Video Enhancement Nodes/VRGDG LTX MSR Reference Builder
ComfyUI Node

VRGDG LTX MSR Reference Builder

Build the multi-subject reference sheet in-node, not in a paint app

By vrgamegirl19·Created about a year ago·Updated about 7 hours ago· 718
VRGDG LTX MSR Reference Builder
    • output
    subject_1
    subject_2
    subject_3
    subject_4
    background_image
    background_modeuse_uploaded_background
    width736
    height1280
    reference_strengthauto - based on subject count

    The LTX 2.3 MSR-LoRA is the community's favorite answer to the model's worst problem: a character can't hold identity across shots. MSR (multiple-subject reference) conditions on several reference images plus a background, and the community that uses it reports real continuity wins on hardware a 4090 owner would snort at - 940x704 in a few minutes on a 4070 with 64GB of RAM. But every MSR workflow needs a reference sheet, and building one by hand in an image editor is the annoying part. VRGDG LTX MSR Reference Builder is the node that assembles that sheet for you, inside the graph.

    How it works

    You pick up to four subject images (subject_1 through subject_4) plus a background_image, and the node lays them out into a single reference composite at the width/height you set - the default is 736x1280, i.e. portrait, which matches how these reference sheets are typically fed to the model. background_mode has two states: use_uploaded_background, which keeps your real background in the sheet, or neutral_placeholder_wip, which swaps in a neutral placeholder - that's the "I'm still sketching this scene, don't pin the background yet" mode.

    The interesting control is reference_strength. Rather than forcing you to guess, it defaults to auto - based on subject count (more subjects, weaker each) and otherwise offers explicit presets: 17 - light, 25 - balanced, 33 - strong, 41 - strongest. These are percentages that decide how hard the reference conditioning leans on the sheet, and they map directly to the LoRA strength discipline the MSR community preaches - too much and subjects bleed into each other, too little and the model ignores them. Start at auto or balanced; only push toward the strong end when a subject keeps drifting.

    The output is a single IMAGE - the assembled sheet - which then feeds the MSR-LoRA's reference input in your conditioning chain.

    The inputs that matter

    • subject_1subject_4 - up to four character/subject images. Unused ones should be (none).
    • background_image - the scene background, or leave it neutral while you work out the characters.
    • width / height - sheet dimensions; keep them divisible by 32 for LTX 2.3.
    • reference_strength - auto by subject count, or the explicit 17/25/33/41 presets.

    Install

    Pack-wide install: ComfyUI Manager → search vrgamedev (or add https://github.com/vrgamegirl19/comfyui-vrgamedevgirl), restart, hard-refresh. Manual: git clone https://github.com/vrgamegirl19/comfyui-vrgamedevgirl into custom_nodes, then pip install -r requirements.txt; on Windows portable, install Cython and scikit-build-core first and prefer Python 3.12.

    Troubleshooting

    If your sheet renders with a placeholder background when you uploaded a real one, check background_mode - it's a two-state switch, and neutral_placeholder_wip is the silent gotcha. If one character keeps turning into another, that's over-conditioning: drop reference_strength a step before you go re-crop the source images. And if nothing's in the sheet at all, the subject enums are file pickers - make sure you've actually chosen images rather than leaving everything at (none).

    CategoryVRGDG/LTX MSR

    Inputs (9)

    NameTypeDefaultDescription
    subject_1COMBO3 options: (none), example.png, vrgdg_placeholder_i2i.png
    subject_2COMBO3 options: (none), example.png, vrgdg_placeholder_i2i.png
    subject_3COMBO3 options: (none), example.png, vrgdg_placeholder_i2i.png
    subject_4COMBO3 options: (none), example.png, vrgdg_placeholder_i2i.png
    background_imageCOMBO3 options: (none), example.png, vrgdg_placeholder_i2i.png
    background_modeCOMBOuse_uploaded_background2 options: use_uploaded_background, neutral_placeholder_wip
    widthINT73632–8192
    heightINT128032–8192
    reference_strengthCOMBOauto - based on subject count5 options: auto - based on subject count, 17 - light, 25 - balanced, 33 - strong, 41 - strongest

    Outputs (1)

    NameTypeDescription
    outputIMAGE