Nodes/Nougan Nodes/Nougan Ultimate Model Loader πŸš€
ComfyUI Node

Nougan Ultimate Model Loader πŸš€

Loads any diffusers/UNET checkpoint (bf16, int8 convrot, w6a8/w4a8 convrot, and new w*a8 formats as ComfyUI adds them), applies attention/efficiency backends, and loads up to two CLIPs plus video/default and audio VAEs.

By WinnouganΒ·Created 4 months agoΒ·Updated 3 days agoΒ· 31
Nougan Ultimate Model Loader πŸš€
    • MODEL
    • CLIP (main)
    • CLIP (enhancer)
    • VAE (video/default)
    • VAE (audio)
    β—„model_nameβ–Ύβ–Ί
    β—„quant_formatautoβ–Ί
    β—„model_typedefaultβ–Ί
    β—„kitchentrueβ–Ί
    β—„sage2falseβ–Ί
    β—„sage3falseβ–Ί
    β—„tritonfalseβ–Ί
    β—„flash2falseβ–Ί
    β—„pytorchfalseβ–Ί
    β—„sparsefalseβ–Ί
    β—„vsafalseβ–Ί
    β—„clip1_nameβ–Ύβ–Ί
    β—„clip1_typeautoβ–Ί
    β—„clip1_devicedefaultβ–Ί
    β—„clip2_enabledfalseβ–Ί
    β—„clip2_nameβ–Ύβ–Ί
    β—„clip2_typeautoβ–Ί
    β—„clip2_devicedefaultβ–Ί
    β—„vae_nameβ–Ύβ–Ί
    β—„audio_vae_enabledfalseβ–Ί
    β—„audio_vae_nameβ–Ύβ–Ί
    β—„heading_modelModelβ–Ί
    β—„heading_engineEfficienciesβ–Ί
    β—„heading_clipsText Encodersβ–Ί
    β—„heading_vaesVAEsβ–Ί
    β—„section_ordermodel,engine,clips,vaesβ–Ί
    CategoryNougan/Loaders

    Inputs (26)

    NameTypeDefaultDescription
    model_nameCOMBODiffusion model / UNET from the diffusers (diffusion_models) folder.
    quant_formatCOMBOautoExpected weight format. 'auto' trusts the checkpoint's embedded quant metadata (int8 / w6a8 / w4a8 convrot etc. are detected automatically). Pick a format to validate it after load β€” you'll get a warning if the file doesn't match.
    model_typeCOMBOdefaultPreset family β€” drives 'auto' CLIP types and guards (VSA needs minimax_h3).
    kitchenBOOLEANtruecomfy-kitchen INT8 attention (dense backend candidate).
    sage2BOOLEANfalseSageAttention 2 (dense backend candidate). Beats comfy-kitchen in priority.
    sage3BOOLEANfalseSageAttention 3 (Blackwell; dense backend candidate). Highest priority.
    tritonBOOLEANfalseOpt into ComfyUI's triton backend for comfy-kitchen quant ops (global session flag, like --enable-triton-backend).
    flash2BOOLEANfalseFlashAttention 2 (dense backend candidate).
    pytorchBOOLEANfalseForce PyTorch SDPA attention explicitly (dense backend candidate).
    sparseBOOLEANfalseBlock-sparse attention (sol-attn) on eligible layers; wraps the dense backend. Long sequences only β€” short ones stay dense.
    vsaBOOLEANfalseVideo Sparse Attention (FastH3) β€” MiniMax-H3 only; conflicts with sparseattention.
    clip1_nameCOMBOMain text encoder (from text_encoders folder).
    clip1_typeCOMBOauto'auto' infers from the model type selector.
    clip1_deviceCOMBOdefault'cpu' pins this encoder to system RAM.
    clip2_enabledBOOLEANfalseLoad a second CLIP (e.g. for a prompt enhancer chain).
    clip2_nameCOMBOSecond text encoder β€” used when its switch is on.
    clip2_typeCOMBOauto36 options: auto, stable_diffusion, sd3, flux, hunyuan_video, wan, +30
    clip2_deviceCOMBOdefault2 options: default, cpu
    vae_nameCOMBOVideo/default VAE: the image VAE for image models, the video VAE for MiniMax-H3.
    audio_vae_enabledBOOLEANfalseAlso load an audio VAE (MiniMax-H3 audio, LTX-2, ...).
    audio_vae_nameCOMBOAudio VAE file β€” used when its switch is on.
    heading_modelSTRINGModelβ€”
    heading_engineSTRINGEfficienciesβ€”
    heading_clipsSTRINGText Encodersβ€”
    heading_vaesSTRINGVAEsβ€”
    section_orderSTRINGmodel,engine,clips,vaesβ€”

    Outputs (5)

    NameTypeDescription
    MODELMODELDiffusion model, patched with the chosen efficiency backends.
    CLIP (main)CLIPMain text encoder for the diffusion model.
    CLIP (enhancer)CLIPSecond text encoder (e.g. prompt enhancer) β€” None while its switch is off.
    VAE (video/default)VAEVideo/default VAE (image VAE for image models, video VAE for MiniMax-H3 etc.).
    VAE (audio)VAEAudio VAE β€” None while its switch is off.