README.md

SVDQuant ComfyUI Node

Installation

Please first install nunchaku following the instructions in README.md.

ComfyUI-CLI

comfy node registry-install svdquant

ComfyUI-Manager (Experimental)

1. Install ComfyUI-Manager with the following commands then restart ComfyUI:

   cd ComfyUI/custom_nodes
   git clone https://github.com/ltdrdata/ComfyUI-Manager comfyui-manager

2. Open the Manager, search svdquant in the Custom Nodes Manager and then install it.

Manual Installation

1. Install dependencies needed to run custom ComfyUI nodes:

   pip install git+https://github.com/asomoza/image_gen_aux.git

2. Set up the dependencies for ComfyUI with the following commands:

   git clone https://github.com/comfyanonymous/ComfyUI.git
   cd ComfyUI
   pip install -r requirements.txt

3. Navigate to the root directory of ComfyUI and link (or copy) the nunchaku/comfyui folder to custom_nodes/svdquant. For example:

   # Clone repositories (skip if already cloned)
   git clone https://github.com/comfyanonymous/ComfyUI.git
   git clone https://github.com/mit-han-lab/nunchaku.git
   cd ComfyUI
   
   # Add SVDQuant nodes
   cd custom_nodes
   ln -s ../../nunchaku/comfyui svdquant

Usage

1. Set Up ComfyUI and SVDQuant:

   • SVDQuant workflows can be found at workflows. You can place them in user/default/workflows in ComfyUI root directory to load them. For example:

     cd ComfyUI
     
     # Copy workflow configurations
     mkdir -p user/default/workflows
     cp ../nunchaku/comfyui/workflows/* user/default/workflows/

   • Install missing nodes (e.g., comfyui-inpainteasy) following this tutorial.

2. Download Required Models: Follow this tutorial and download the required models into the appropriate directories using the commands below:

   huggingface-cli download comfyanonymous/flux_text_encoders clip_l.safetensors --local-dir models/text_encoders
   huggingface-cli download comfyanonymous/flux_text_encoders t5xxl_fp16.safetensors --local-dir models/text_encoders
   huggingface-cli download black-forest-labs/FLUX.1-schnell ae.safetensors --local-dir models/vae

3. Run ComfyUI: From ComfyUI’s root directory, execute the following command to start the application:

   python main.py

4. Select the SVDQuant Workflow: Choose one of the SVDQuant workflows (workflows that start with svdq-) to get started. For the FLUX.1-Fill workflow, you can use the built-in MaskEditor tool to add mask on top of an image.

SVDQuant Nodes

• SVDQuant Flux DiT Loader: A node for loading the FLUX diffusion model.

  • model_path: Specifies the model location. If set to mit-han-lab/svdq-int4-flux.1-schnell, mit-han-lab/svdq-int4-flux.1-dev, mit-han-lab/svdq-int4-flux.1-canny-dev, mit-han-lab/svdq-int4-flux.1-fill-dev or mit-han-lab/svdq-int4-flux.1-depth-dev, the model will be automatically downloaded from our Hugging Face repository. Alternatively, you can manually download the model directory by running the following command example:

    huggingface-cli download mit-han-lab/svdq-int4-flux.1-dev --local-dir models/diffusion_models/svdq-int4-flux.1-dev

    After downloading, specify the corresponding folder name as the model_path.

  • device_id: Indicates the GPU ID for running the model.

• SVDQuant FLUX LoRA Loader: A node for loading LoRA modules for SVDQuant FLUX models.

  • Place your LoRA checkpoints in the models/loras directory. These will appear as selectable options under lora_name. Meanwhile, the example Ghibsky LoRA is included and will automatically download from our Hugging Face repository when used.

  • lora_format specifies the LoRA format. Supported formats include:

    • diffusers (e.g., aleksa-codes/flux-ghibsky-illustration)
    • comfyui (e.g., Shakker-Labs/FLUX.1-dev-LoRA-Children-Simple-Sketch)
    • xlab (e.g., XLabs-AI/flux-RealismLora)
    • svdquant (e.g., mit-han-lab/svdquant-lora-collection).

  • base_model_name specifies the path to the quantized base model. If lora_format is already set to svdquant, this option has no use. You can set it to the same value as model_path in the above SVDQuant Flux DiT Loader.

  • Note: Currently, only one LoRA can be loaded at a time.

• SVDQuant Text Encoder Loader: A node for loading the text encoders.

  • For FLUX, use the following files:

    • text_encoder1: t5xxl_fp16.safetensors
    • text_encoder2: clip_l.safetensors

  • t5_min_length: Sets the minimum sequence length for T5 text embeddings. The default in DualCLIPLoader is hardcoded to 256, but for better image quality in SVDQuant, use 512 here.

  • t5_precision: Specifies the precision of the T5 text encoder. Choose INT4 to use the INT4 text encoder, which reduces GPU memory usage by approximately 15GB. Please install deepcompressor when using it:

    git clone https://github.com/mit-han-lab/deepcompressor
    cd deepcompressor
    pip install poetry
    poetry install

    • int4_model: Specifies the INT4 model location. This option is only used when t5_precision is set to INT4. By default, the path is mit-han-lab/svdq-flux.1-t5, and the model will automatically download from our Hugging Face repository. Alternatively, you can manually download the model directory by running the following command:

      huggingface-cli download mit-han-lab/svdq-flux.1-t5 --local-dir models/text_encoders/svdq-flux.1-t5

      After downloading, specify the corresponding folder name as the int4_model.

• FLUX.1 Depth Preprocessor: A node for loading the depth estimation model and output the depth map. model_path specifies the model location. If set to LiheYoung/depth-anything-large-hf, the model will be automatically downloaded from the Hugging Face repository. Alternatively, you can manually download the repository at models/checkpoints by running the following command example:

  huggingface-cli download LiheYoung/depth-anything-large-hf --local-dir models/checkpoints/depth-anything-large-hf