A collection of INT4 ConvRot-quantized diffusion, video, and upscaling models for ComfyUI, built to run comfortably on 8GB-class GPUs (developed and tested on an RTX 3070 Ti) without gutting output quality.
Standard INT8/INT4 row-wise quantization throws away a lot of precision on the weight matrices that matter most for visual fidelity. ConvRot groups weights along their largest power-of-4-compatible dimension before quantizing, which keeps much more of the original model's detail and reduces the artifacting you'd normally see from a naive INT4 cast. The trade-off is VRAM and disk savings big enough to run models like SeedVR2 7B and full video diffusion checkpoints on 8GB cards.
Requirements
ComfyUI (nighlty build)
If you're getting chronic errors update your Conda environment (I'm running Pytorch 2.12, cu132, Python 3.12, Flashattention/Sageattention and Triton 3.8)
Installation
Install ComfyUI-INT4-Fast into ComfyUI/custom_nodes/
Download the model(s) you want from this repo into the matching ComfyUI/models/diffusion_models/ (or appropriate folder for video/upscale models)
Load with the INT4 loader node from ComfyUI-INT4-Fast — do not use the standard checkpoint/UNETLoader nodes, they will not decode these correctly
See the Samples and Workflow folder in this repo for ready-to-use ComfyUI workflow JSONs and sample outputs
Quantization pipeline
Built with Starnodes power:
Grab the Starnodes model converter and do it yourself if you wish. It supports INT8 and INT4 convrot:
Starnodes
Links
🎥 YouTube: tutorials and walkthroughs for this collection
💬 Discord: community, support, and early access
🩷 Patreon / ☕ Ko-fi: support ongoing quantization work
Inherits the license terms of each respective base model (Krea 2, LTX-2.3, Sulphur 2, SeedVR2). Check each upstream model's license before commercial use.
