Nunchaku
A high-performance inference engine for 4-bit diffusion models, based on the SVDQuant quantization method, supporting FLUX and Qwen-Image models with ComfyUI integration.
Nunchaku is an inference engine for 4-bit neural networks, built around SVDQuant, a technique that absorbs outlier values into low-rank components to keep 4-bit diffusion models accurate. Its quantization step is implemented in a companion library called DeepCompressor.
It supports FLUX and Qwen-Image family diffusion models, including Lightning and Turbo variants, and offers a Python backend, native ComfyUI nodes with LoRA support, and asynchronous offloading that can reduce VRAM use for large models to a few gigabytes. Recent releases have added INT4 support for older 20-series GPUs and compatibility with newer ComfyUI versions.
Its methods were introduced in a paper accepted as an ICLR 2025 Spotlight.