ComfyUI-GGUF 🧩 Workflow Open source

What it is

ComfyUI-GGUF by city96 (Apache-2.0) adds loaders for models stored in the GGUF quantised format, the same format popularised by llama.cpp. Quantised versions of large image and video diffusion models (FLUX, Wan, HunyuanVideo, Qwen-Image and others) plus their T5 text encoders take much less VRAM, so they fit on consumer GPUs with a modest quality cost. Community uploaders publish GGUF versions of most major open models soon after release, usually at several quantisation levels (for example Q4, Q5, Q8), so you can trade quality for memory to suit your card.

Key features

• Unet/diffusion model loader for GGUF files
• GGUF CLIP/T5 text-encoder loaders
• Works with transformer-based models such as FLUX, Wan and HunyuanVideo
• Wide choice of quantisation levels from community uploads
• Tools for converting your own models to GGUF
• Drop-in replacement for standard loaders in existing workflows

How to install

Run git clone https://github.com/city96/ComfyUI-GGUF ComfyUI/custom_nodes/ComfyUI-GGUF and install the gguf Python package, or install through ComfyUI-Manager. Put .gguf model files in ComfyUI/models/unet.

Pricing and rights

Free, Apache-2.0. Quantised models inherit the original model's licence (for example FLUX.1 [dev] remains non-commercial).

Who it is for

ComfyUI users on 8-16 GB GPUs, or laptops, who want to run large modern image and video models without renting cloud GPUs.

Verdict

A small, reliable pack that makes big models practical on consumer hardware and is used in many low-VRAM tutorials. Updates are infrequent (last push January 2026), and heavy quantisation visibly reduces quality.