Extensions/ComfyUI-WeeLLM
ComfyUI Extension

ComfyUI-WeeLLM

Native ComfyUI integration for WeeLLM: Run multi-billion parameter diffusion models (FLUX, SD3, MiniMax) under 4GB VRAM using layer-streaming inference.

By Jit-Roy·Created 2 months ago·Updated 7 days ago· 7
Jit-Roy/ComfyUI-WeeLLM
Nodes3
On cloudLocal install
CategoryWeeLLM
Stars7
Updated7 days ago

Nodes (3)

WeeLLM Generate

Text-to-image, img2img, and inpainting in one node

WeeLLM
WeeLLM Loader

A 12B FLUX pipeline on a 4 GB card, no quantization

WeeLLM
WeeLLM Video Generate

MiniMax-H3 clips on a GPU with nothing to spare

WeeLLM
Readme
<div align="center"> <img src="docs/banner.png" alt="ComfyUI-WeeLLM" width="700"/> <br /> <a href="https://camo.githubusercontent.com/5e3ae50ed74ba1fea6f7b2002fe729b6dac58900dde6ba310eaa26b56dd071dd/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f5374617475732d4163746976652d737563636573732e737667"><img src="https://camo.githubusercontent.com/5e3ae50ed74ba1fea6f7b2002fe729b6dac58900dde6ba310eaa26b56dd071dd/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f5374617475732d4163746976652d737563636573732e737667" alt="Status: Active" /></a> <a href="https://camo.githubusercontent.com/7013272bd27ece47364536a221edb554cd69683b68a46fc0ee96881174c4214c/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f6c6963656e73652d4d49542d626c75652e737667"><img src="https://camo.githubusercontent.com/7013272bd27ece47364536a221edb554cd69683b68a46fc0ee96881174c4214c/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f6c6963656e73652d4d49542d626c75652e737667" alt="License: MIT" /></a> <a href="https://camo.githubusercontent.com/dd0b24c1e6776719edb2c273548a510d6490d8d25269a043dfabbd38419905da/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f5052732d77656c636f6d652d627269676874677265656e2e737667"><img src="https://camo.githubusercontent.com/dd0b24c1e6776719edb2c273548a510d6490d8d25269a043dfabbd38419905da/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f5052732d77656c636f6d652d627269676874677265656e2e737667" alt="PRs Welcome" /></a> </div>

A native ComfyUI custom node wrapper for WeeLLM — bringing ultra-low VRAM layer-streaming inference directly into your Comfy workflows.

With ComfyUI-WeeLLM, you can run massive diffusion models (like FLUX, Stable Diffusion, Stable Diffusion XL, Z image, Krea2, Qwen Image, Ltx2.5 and MiniMax-H3) on GPUs with less than 4GB of VRAM, without requiring to quantize.

  • GGUFs and Safetensors - Everything is Supported.
  • LoRa Support Available.

Workflows

Workflow 1: Image to Image

This workflow uses an input image as the starting point and generates a revised image according to the prompt and strength value.

Workflow 1: Image to Image

Download the Image to Image workflow

  1. Import the workflow into ComfyUI and load an input image.
  2. In WeeLLM Loader, set model_path to a supported Hugging Face model ID or local model directory and set task to image-to-image.
  3. Connect the input image to the image input on WeeLLM Generate.
  4. Adjust strength, then queue the prompt.

Workflow 2: Image to Video

This workflow uses an input image as the first frame of a generated video. The resulting frame batch can be sent to a Video Combine node.

Workflow 2: Image to Video

Download the Image to Video workflow

  1. Import the workflow into ComfyUI and load an input image.
  2. In WeeLLM Loader, set task to video.
  3. Connect the input image to the first_frame input on WeeLLM Video Generate.
  4. Connect the generated frames to a Video Combine node, then queue the prompt.

Installation

Method 1: ComfyUI Manager (Recommended)

  1. Open the ComfyUI Manager.
  2. Click Install Custom Nodes.
  3. Search for ComfyUI-WeeLLM and click Install.
  4. Restart ComfyUI.

Method 2: Manual Git Clone

Navigate to your ComfyUI/custom_nodes/ directory and run:

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

Nodes

1. WeeLLM Loader

Responsible for configuring and caching the model pipeline.

  • model_path: Provide the Hugging Face repo ID (e.g. black-forest-labs/FLUX.1-schnell) or a local directory.
  • task: Select the type of pipeline to initialize (text-to-image, image-to-image, video).
  • dtype: The precision to use (e.g., bfloat16).

2. WeeLLM Generate

Generates images using the loaded pipeline.

  • Connect the WEE_PIPELINE from the Loader to this node.
  • Supports optional image and mask_image inputs for automatic Image-to-Image and Inpainting.

3. WeeLLM Video Generate

Generates video batches for models like MiniMax-H3 FL2VA.

  • Connect the WEE_PIPELINE from the Loader (with video task selected) to this node.
  • Outputs an IMAGE batch that can be sent directly to Video Combine nodes.