Nodes/ComfyUI_SmartModelLoader/Model Loader
ComfyUI Node

Model Loader

Standalone model loader with direct model/clip/vae outputs. Supports checkpoints, UNet, Nunchaku, and GGUF with LoRA and BlockSwap.

By r-vage·Created 9 days ago·Updated 2 days ago· 2
Model Loader
    • model
    • clip
    • vae
    • audio_vae
    • model_name
    ◄featuresmemory_cleanup►
    ◄model_typeStandard Checkpoint►
    ◄ckpt_nameNone►
    ◄unet_nameNone►
    ◄nunchaku_nameNone►
    ◄qwen_nameNone►
    ◄zimage_nameNone►
    ◄gguf_nameNone►
    ◄weight_dtypedefault►
    ◄data_typebfloat16►
    ◄cache_threshold0.000►
    ◄attentionflash-attention2►
    ◄i2f_modeenabled►
    ◄cpu_offloadauto►
    ◄num_blocks_on_gpu30►
    ◄use_pin_memoryenable►
    ◄gguf_dequant_dtypedefault►
    ◄gguf_patch_dtypedefault►
    ◄gguf_patch_on_devicefalse►
    ◄enable_clip_layertrue►
    ◄stop_at_clip_layer-2►
    ◄lora_count1►
    ◄lora_switch_1false►
    ◄lora_name_1None►
    ◄lora_weight_11.00►
    ◄lora_switch_2false►
    ◄lora_name_2None►
    ◄lora_weight_21.00►
    ◄lora_switch_3false►
    ◄lora_name_3None►
    ◄lora_weight_31.00►
    ◄sampling_methodNone►
    ◄sampling_subtypeeps►
    ◄shift3.00►
    ◄base_shift0.50►
    ◄sampling_width1024►
    ◄sampling_height1024►
    ◄original_timesteps50►
    ◄zsnrfalse►
    ◄sigma_max120.000►
    ◄sigma_min0.002►
    ◄blocks_to_swap5►
    ◄offload_embeddingsfalse►
    ◄ltx_text_encoderNone►
    Category🌒 Smart Model Loader/ Loader

    Inputs (44)

    NameTypeDefaultDescription
    featuresSTRINGmemory_cleanupComma-separated feature list. JS combo-chip replaces this widget.
    model_typeCOMBOStandard CheckpointSelect model format
    ckpt_nameCOMBONoneSelect checkpoint file
    unet_nameCOMBONoneSelect UNet diffusion model
    nunchaku_nameCOMBONoneSelect Nunchaku Flux model
    qwen_nameCOMBONoneSelect Nunchaku Qwen model
    zimage_nameCOMBONoneSelect Nunchaku ZImage model
    gguf_nameCOMBONoneSelect GGUF model
    weight_dtypeCOMBOdefaultWeight dtype for UNet model
    data_typeCOMBObfloat16Model data type for Nunchaku
    cache_thresholdFLOAT0.0000–1Cache threshold for Nunchaku
    attentionCOMBOflash-attention2Attention implementation
    i2f_modeCOMBOenabledGEMM implementation
    cpu_offloadCOMBOautoCPU offload
    num_blocks_on_gpuINT301–60Blocks on GPU (Nunchaku Qwen/ZImage)
    use_pin_memoryCOMBOenableUse pinned memory
    gguf_dequant_dtypeCOMBOdefaultDequantization dtype
    gguf_patch_dtypeCOMBOdefaultLoRA patch dtype
    gguf_patch_on_deviceBOOLEANfalseApply patches on GPU
    enable_clip_layerBOOLEANtrueTrim baked CLIP to specific layer (Standard Checkpoint only)
    stop_at_clip_layerINT-2-24–-1CLIP layer to stop at
    lora_countCOMBO1Number of LoRA slots
    lora_switch_1BOOLEANfalseEnable LoRA 1
    lora_name_1COMBONoneLoRA 1 file
    lora_weight_1FLOAT1.00-10–10LoRA 1 model weight
    lora_switch_2BOOLEANfalseEnable LoRA 2
    lora_name_2COMBONoneLoRA 2 file
    lora_weight_2FLOAT1.00-10–10LoRA 2 model weight
    lora_switch_3BOOLEANfalseEnable LoRA 3
    lora_name_3COMBONoneLoRA 3 file
    lora_weight_3FLOAT1.00-10–10LoRA 3 model weight
    sampling_methodCOMBONoneSampling method: SD3 (shift=3.0), AuraFlow (shift=1.73), Flux (max_shift=1.15), Stable Cascade (shift=2.0), LCM (distilled), ContinuousEDM/V (continuous sampling), LTXV (video)
    sampling_subtypeCOMBOepsSubtype for ContinuousEDM sampling
    shiftFLOAT3.000–100Universal shift parameter (SD3: 3.0, AuraFlow: 1.73, Flux max_shift: 1.15, Stable Cascade: 2.0)
    base_shiftFLOAT0.500–100Base shift for Flux/LTXV sampling (default: 0.5)
    sampling_widthINT102416–32768Width for Flux sampling shift calculation
    sampling_heightINT102416–32768Height for Flux sampling shift calculation
    original_timestepsINT501–1000Original timesteps for LCM sampling (default: 50)
    zsnrBOOLEANfalseEnable zero-terminal SNR for LCM sampling
    sigma_maxFLOAT120.0000–1000Maximum sigma for ContinuousEDM/V sampling (EDM: 120.0, V: 500.0)
    sigma_minFLOAT0.0020–1000Minimum sigma for ContinuousEDM/V sampling (EDM: 0.002, V: 0.03)
    blocks_to_swapINT50–100Number of transformer blocks to offload from GPU to CPU. Higher = more VRAM saved but slower inference. Suggested ~value (max total blocks): flux/chroma ~10 (max 57), sd3 ~8 (max 24-38), wan ~10 (max 30-40), hunyuan-video ~10 (max 60), ltxv ~6 (max 28), cosmos ~8 (max 28-36), zimage ~10 (max 30), qwenimage ~20 (max 60), mochi ~10 (max 48), hidream ~10 (max 48). Set to 0 to disable.
    offload_embeddingsBOOLEANfalseAlso offload embedding and projection layers for extra VRAM savings.
    ltx_text_encoderCOMBONoneOptional LTX2/LTXV gemma text encoder (from the text_encoders/clip folder, GGUF or safetensors). When set, it is combined with the loaded Standard Checkpoint / UNet file's baked text-projection to build a correct LTXAV CLIP, overriding the (empty) baked CLIP. Leave as None for normal baked-CLIP behavior.

    Outputs (5)

    NameTypeDescription
    modelMODEL—
    clipCLIP—
    vaeVAE—
    audio_vaeVAE—
    model_nameSTRING—