Nodes/ComfyUI_SmartModelLoader/Smart Model Loader
ComfyUI Node

Smart Model Loader

All-in-one model loader with multi-select feature toggling. Supports Standard Checkpoints, UNet, Nunchaku (Flux/Qwen/ZImage), and GGUF models.

By r-vage·Created 9 days ago·Updated 2 days ago· 2
Smart Model Loader
    • pipe
    ◄featuresclip,vae,memory_cleanup►
    ◄template_actionNone►
    ◄template_nameNone►
    ◄new_template_name►
    ◄model_typeStandard Checkpoint►
    ◄ckpt_nameNone►
    ◄unet_nameNone►
    ◄nunchaku_nameNone►
    ◄qwen_nameNone►
    ◄zimage_nameNone►
    ◄gguf_nameNone►
    ◄weight_dtypedefault►
    ◄data_typebfloat16►
    ◄cache_threshold0.0►
    ◄attentionflash-attention2►
    ◄i2f_modeenabled►
    ◄cpu_offloadauto►
    ◄num_blocks_on_gpu30►
    ◄use_pin_memoryenable►
    ◄gguf_dequant_dtypedefault►
    ◄gguf_patch_dtypedefault►
    ◄gguf_patch_on_devicefalse►
    ◄blocks_to_swap10►
    ◄offload_embeddingsfalse►
    ◄sampling_methodNone►
    ◄sampling_subtypeeps►
    ◄shift3.0►
    ◄base_shift0.5►
    ◄sampling_width1024►
    ◄sampling_height1024►
    ◄original_timesteps50►
    ◄zsnrfalse►
    ◄sigma_max120.00►
    ◄sigma_min0.00►
    ◄clip_sourceBaked►
    ◄clip_count1►
    ◄clip_name1None►
    ◄clip_name2None►
    ◄clip_name3None►
    ◄clip_name4None►
    ◄clip_typeflux►
    ◄enable_clip_layertrue►
    ◄stop_at_clip_layer-2►
    ◄vae_sourceBaked►
    ◄vae_nameNone►
    ◄resolution1024x1024 (1:1 XL/SD3/Flux/HiDream)►
    ◄width1024►
    ◄height1024►
    ◄lora_count1►
    ◄lora_switch_1false►
    ◄lora_name_1None►
    ◄lora_weight_11.0►
    ◄lora_switch_2false►
    ◄lora_name_2None►
    ◄lora_weight_21.0►
    ◄lora_switch_3false►
    ◄lora_name_3None►
    ◄lora_weight_31.0►
    ◄sampler_nameeuler►
    ◄schedulernormal►
    ◄steps20►
    ◄cfg8.0►
    ◄denoise1.00►
    ◄flux_guidance3.5►
    ◄batch_size1►
    ◄audio_vae_sourceExternal►
    ◄audio_vae_nameNone►
    ◄verify_fileoff►
    ◄expected_hashes{}►
    ◄air_or_hash►
    ◄download_locators[]►
    ◄download_target_role►
    ◄model_precisiondefault►
    ◄seed0►
    Category🌒 Smart Model Loader/ Loader

    Inputs (74)

    NameTypeDefaultDescription
    featuresSTRINGclip,vae,memory_cleanupComma-separated list of enabled loader features. Selected via the Mode Bar chip widget on the frontend.
    template_actionCOMBONoneManage saved presets/templates for this node: • None: Do nothing • Load: Load configurations from template_name • Save: Save current configurations into a new template.
    template_nameCOMBONoneSelect a saved configuration template to load or delete.
    new_template_nameSTRINGEnter a name for the new configuration template when saving current settings.
    model_typeCOMBOStandard CheckpointType of model architecture to load: • Standard Checkpoint: Full model (diffusion, CLIP, VAE) in one file • UNet Model: Diffusion-only model • Nunchaku: FP4 quantized GPU inference models • GGUF Model: llama.cpp/GGUF quantized models
    ckpt_nameCOMBONoneSelect a standard Stable Diffusion or Flux checkpoint file containing diffusion, CLIP, and VAE weights.
    unet_nameCOMBONoneSelect a standalone UNet/Diffusion model checkpoint (e.g. Flux, SD3, AuraFlow) from the diffusion_models directory.
    nunchaku_nameCOMBONoneSelect a Nunchaku FP4-quantized model for Flux to load.
    qwen_nameCOMBONoneSelect a Nunchaku FP4-quantized model for Qwen2-VL to load.
    zimage_nameCOMBONoneSelect a Nunchaku FP4-quantized model for ZImage to load.
    gguf_nameCOMBONoneSelect a GGUF format diffusion model from the diffusion_models_gguf directory.
    weight_dtypeCOMBOdefaultPreferred weight precision for loading checkpoint weights (e.g. default, fp8_e4m3fn, fp8_e5m2, bfloat16). FP8 saves VRAM.
    data_typeCOMBObfloat16Data type precision for Nunchaku FP4 model layers. Select bfloat16 or float16.
    cache_thresholdFLOAT0.00–1GPU memory caching threshold for Nunchaku FP4 layers. Higher values reserve more GPU cache for layers.
    attentionCOMBOflash-attention2Attention implementation variant: • flash-attention2: Highly optimized for modern Ampere/Ada GPUs • nunchaku-fp16: Optimized FP16 attention kernel
    i2f_modeCOMBOenabledGEMM matrix multiplication implementation mode for Nunchaku GPU execution.
    cpu_offloadCOMBOautoToggle offloading of inactive parts of the model from VRAM to CPU RAM to save VRAM.
    num_blocks_on_gpuINT301–60Number of transformer blocks to keep on the GPU for Nunchaku Qwen. Remaining blocks are offloaded to CPU.
    use_pin_memoryCOMBOenableEnable pinned memory (host-allocated memory) for faster tensor transfers between system RAM and GPU VRAM.
    gguf_dequant_dtypeCOMBOdefaultDequantization data type precision for GGUF weights when patching them (default, float16, bfloat16, float32).
    gguf_patch_dtypeCOMBOdefaultData type precision to use when applying LoRA patches to GGUF weights.
    gguf_patch_on_deviceBOOLEANfalseApply patches directly on the GPU rather than host CPU memory (faster but requires more VRAM during load).
    blocks_to_swapINT100–100Number of transformer blocks to swap/offload from GPU to CPU memory. Higher values save significant VRAM but slow down generation. Recommended values (max blocks): • Flux: ~10 (max 57) • SD3: ~8 (max 24) • Wan2.1: ~10 (max 40) • HunyuanVideo: ~10 (max 60) • LTX-Video: ~6 (max 28) Set to 0 to disable.
    offload_embeddingsBOOLEANfalseOffloads the embedding and projection layers (text_embedding, img_emb, time_in) to CPU RAM. Saves ~100-300MB VRAM at a slight speed cost.
    sampling_methodCOMBONoneModel-level sampling correction: • SD3/AuraFlow/Flux/Stable Cascade: Sets standard scheduler shifts • LCM: Configures latent consistency model scheduling • ContinuousEDM/ContinuousV: EDM scheduler • LTXV: LTX-Video scheduler
    sampling_subtypeCOMBOepsSubtype scheduling curve for ContinuousEDM sampling (e.g. eps, v_prediction, edm, cosmos_rflow).
    shiftFLOAT3.00–10Universal scheduling shift multiplier. SD3 default: 3.0, AuraFlow: 1.73, Stable Cascade: 2.0.
    base_shiftFLOAT0.50–10Base scheduling shift for Flux (default: 0.5) and LTX-Video (default: 2.05).
    sampling_widthINT102416–2000Target resolution width for Flux/LTXV sampling shift calculations. Used to scale scheduler step sizes.
    sampling_heightINT102416–2000Target resolution height for Flux/LTXV sampling shift calculations. Used to scale scheduler step sizes.
    original_timestepsINT501–1000Original training timesteps of the LCM model (used to scale distilled step sizes).
    zsnrBOOLEANfalseZero-Terminal Signal-to-Noise Ratio (zsnr) adjustment to allow generating true darks/blacks.
    sigma_maxFLOAT120.000–1000Maximum noise sigma boundary value for ContinuousEDM/ContinuousV scheduling.
    sigma_minFLOAT0.000–1000Minimum noise sigma boundary value for ContinuousEDM/ContinuousV scheduling.
    clip_sourceCOMBOBakedSource of the text encoder (CLIP): • Baked: Uses CLIP embedded in the checkpoint • External: Uses separate CLIP files • External + Model File: Extends external loaders with the UNet file to auto-resolve baked projections (e.g. LTXV Gemma).
    clip_countCOMBO1Number of separate CLIP model/Text Encoder files to load concurrently (e.g., 2 for Flux, 3 for SD3).
    clip_name1COMBONoneSelect the primary CLIP or Text Encoder checkpoint file.
    clip_name2COMBONoneSelect the secondary CLIP or Text Encoder checkpoint file.
    clip_name3COMBONoneSelect the third CLIP or Text Encoder checkpoint file.
    clip_name4COMBONoneSelect the fourth CLIP or Text Encoder checkpoint file.
    clip_typeCOMBOfluxCLIP loader wrapping/architecture mapping to match the target model type (e.g. flux, sd3, sdxl, wan, mochi, ltxv).
    enable_clip_layerBOOLEANtrueEnables stopping CLIP text evaluation at a specific layer (CLIP skip) instead of evaluating all the way to the end.
    stop_at_clip_layerINT-2-24–-1Which layer to stop CLIP text evaluation at (e.g. -2 for SD1.5/SDXL, -3 for SD3). Negative values count back from final layer.
    vae_sourceCOMBOBakedSource of the variational autoencoder (VAE): • Baked: Extract VAE from the checkpoint file • External: Load a standalone VAE file
    vae_nameCOMBONoneSelect a standalone VAE file to load from the vae directory.
    resolutionCOMBO1024x1024 (1:1 XL/SD3/Flux/HiDream)Select a pre-calculated latent size preset (Select 'Custom' to manually specify width and height). Note: Latent properties (channels, downscale ratio) are automatically detected from the active VAE (baked or external).
    widthINT102416–2000Custom width for empty latent generation. Must be a multiple of 8.
    heightINT102416–2000Custom height for empty latent generation. Must be a multiple of 8.
    lora_countCOMBO1Number of active LoRA slots to configure.
    lora_switch_1BOOLEANfalseToggle to quickly enable or disable this LoRA slot without clearing the filename.
    lora_name_1COMBONoneSelect a LoRA model file from the loras directory.
    lora_weight_1FLOAT1.0-10–10Weight scale to apply to this LoRA's weights. 1.0 is standard strength; negative values invert the effect.
    lora_switch_2BOOLEANfalseToggle to quickly enable or disable this LoRA slot without clearing the filename.
    lora_name_2COMBONoneSelect a LoRA model file from the loras directory.
    lora_weight_2FLOAT1.0-10–10Weight scale to apply to this LoRA's weights. 1.0 is standard strength; negative values invert the effect.
    lora_switch_3BOOLEANfalseToggle to quickly enable or disable this LoRA slot without clearing the filename.
    lora_name_3COMBONoneSelect a LoRA model file from the loras directory.
    lora_weight_3FLOAT1.0-10–10Weight scale to apply to this LoRA's weights. 1.0 is standard strength; negative values invert the effect.
    sampler_nameCOMBOeulerSelect the ComfyUI sampling algorithm (e.g. euler, heun, dpmpp_2m).
    schedulerCOMBOnormalSelect the noise scheduling curve (e.g. normal, karras, exponential, sgm_uniform).
    stepsINT201–150Number of denoising steps. Higher values take longer but refine the image; 20-30 steps is standard for most models.
    cfgFLOAT8.01–30Classifier-Free Guidance (CFG) scale. Controls prompt adherence. Higher values enforce the prompt strictly but can burn colors; 1.0 disables it.
    denoiseFLOAT1.000–1Denoising strength. 1.0 fully denoises the input latent; lower values preserve more of the starting image or latent.
    flux_guidanceFLOAT3.50–10Guidance scale specific to Flux models. Controls prompt adherence/contrast without standard CFG burn.
    batch_sizeINT11–4096Number of latent images to generate in parallel in a single execution batch.
    audio_vae_sourceCOMBOExternalSource of the audio decoder/VAE: • External: Loads separate vocoder/audio-VAE files • Baked: Extracts audio VAE directly from LTX2 all-in-one model files
    audio_vae_nameCOMBONoneSelect a standalone LTXV/LTX2 audio VAE file from the vae directory.
    verify_fileCOMBOoffPrimary model integrity mode: • off: No model hashing. • sidecar: Computes and saves a .sha256 baseline next to the selected primary model. • verify: Compares the primary model when a trusted SHA256 is available and stops on mismatch; without one, it records a local baseline and continues loading. External CLIP, VAE, audio VAE, and LoRA files always retain path and safe-format validation. For downloads, paste a CivitAI URN:AIR or SHA256 into air_or_hash.
    expected_hashesSTRING{}Internal JSON database mapping filenames to expected SHA256 and CivitAI AIR metadata. Managed automatically.
    air_or_hashSTRINGPaste CivitAI AIR (urn:air:...) or a SHA256 hash here. Add +<fileId> to an AIR to select an exact file when several artifacts share a precision. Used to verify integrity on load or trigger auto-downloads if files are missing.
    download_locatorsSTRING[]Internal JSON store containing locator downloads. Managed automatically.
    download_target_roleCOMBOSelect which ComfyUI input folder to save downloaded models into (e.g. checkpoints, diffusion_models, vae).
    model_precisionCOMBOdefaultPreferred weight precision (e.g. fp16, bf16, fp8) when downloading models from CivitAI using URN:AIR locators. 'default' grabs the primary file; an explicit precision selects the unique largest matching artifact. Use the AIR +<fileId> suffix when matching sizes are missing or tied.
    seedINT0-3–18446744073709550000Controls generation reproducibility. Use specific values for deterministic output: • -1: Randomize the seed on every execution • -2: Increment the seed by 1 after each run • -3: Decrement the seed by 1 after each run

    Outputs (1)

    NameTypeDescription
    pipePIPE—