GitHub radar
Qwen-Image-2.1 is reduced to 6 steps with Viggle’s quick extension
Viggle has developed a fast augmentation module for Alibaba’s Qwen-Image-2.1, which can complete the image generation and modification task within 6 steps, while maintaining high fidelity.
This is an accelerated plug-in of Alibaba’s Qwen team image model Qwen-Image-2.1. The acceleration model is trained by distillation training to learn how to rapidly reproduce the results of the baseline model. Base model: 40 steps, acceleration model: 6 steps. Support generation (image generation based on text description) and modification (image modification based on 1–3 reference images, including preserving some faces/products, etc.) functions, and keep high similarity with the baseline model in most cases. There might be some differences in some scenarios, like some small texts inside the image or complicated modifications like multi-reference face swap. It requires the original Qwen-Image-2.1 model weight file, so don’t delete the original one.
Why a vibe-coder should care
If you are a user who uses local image generation and waits for 40 steps, now you can save most of the time cost and only lose a little bit of quality in most cases. If you care about some tiny fonts inside the image or multi-reference face swap, please use the slow baseline model. Non-commercial use only, not allowed for commercial use, just for fun or test.
How to install
Copy this and send it to your agent — Claude Code, Codex, any of them:
Set up the fast Viggle version of Qwen Image 2.1: https://huggingface.co/Viggle/Qwen-Image-2.1-viggle-turbo — load it on top of the base Qwen-Image-2.1 model via ComfyUI or diffusers, and show me how to generate an image in six steps instead of forty.
A regular laptop with 16 GB of RAM or more - same as the base Qwen-Image-2.1, about 4 GB in the quantized version. The add-on itself is small and barely adds to that.
Open on Hugging Face▌ More finds