nunchux-ai/ComfyUI-nunchaku

What it solves

This project provides a ComfyUI plugin that allows users to run high-performance, 4-bit quantized neural networks. It solves the problem of high VRAM usage and slow inference speeds when running large image generation models, making them more accessible on consumer-grade hardware, including 20-series GPUs.

How it works

The plugin integrates the Nunchaku inference engine, which uses SVDQuant to quantize models to 4-bit. This reduces the memory footprint and increases speed. It supports various models such as FLUX.1, Qwen-Image, and Z-Image-Turbo, and provides custom ComfyUI nodes for loading these quantized models and managing their execution.

Who it’s for

Digital artists, AI researchers, and ComfyUI users who want to run advanced image generation and editing models with lower VRAM requirements and asynchronous offloading to save memory.

Highlights

• 4-bit Quantization: Uses SVDQuant for efficient 4-bit neural networks.
• Broad Model Support: Compatible with FLUX.1, Qwen-Image, and Z-Image-Turbo.
• VRAM Optimization: Features asynchronous offloading for Qwen-Image, reducing Transformer VRAM usage to as little as 3 GiB.
• Extended Features: Supports LoRA, ControlNet-Union-Pro 2.0, and PuLID integration.
• Hardware Compatibility: Compatible with NVIDIA 20-series GPUs.