Nodes/ComfyUI-GGUF-Loader/Text Encoder + ClipProj Loader ⚡
ComfyUI Node

Text Encoder + ClipProj Loader ⚡

Load a small text encoder and project it into a large one's space, so a 4B or 8B Qwen3-VL can stand in for the 32B MiniMax H3 expects. Matrices go in models/clip_projections/.

By ChrisColeTech·Created 13 days ago·Updated about 3 hours ago· 6
Text Encoder + ClipProj Loader ⚡
    • CLIP
    clip_name
    typeauto
    projection
    Category🤖 CCTech/GGUF

    Inputs (3)

    NameTypeDefaultDescription
    clip_nameCOMBOA full Qwen3-VL text encoder, .gguf or .safetensors. Not an mmproj file, which is the vision projector alone and carries no text model. It must be a VL model: a text-only Qwen3 of the same size loads without complaint and produces conditioning that ignores your prompt.
    typeCOMBOautoauto reads the file header and picks the architecture. Override only if that fails: krea2 = 4B, boogu = 8B, minimax = 32B.
    projectionCOMBOLearned matrix from models/clip_projections/, or a <control:...> baseline. mmh3-4b-* goes with a 4B, mmh3-8b-* with an 8B; they are not interchangeable. Run the controls first: they show what the diffusion model does on its own.

    Outputs (1)

    NameTypeDescription
    CLIPCLIP