The ComfyUI-Spectrum-Qwen2.1 plugin has been released for accelerating image generation

Developers introduced ComfyUI-Spectrum-Qwen2.1 — an extension for the popular generation interface, optimizing the work of the Qwen-Image-2.1 model. The tool allows reducing the request processing time by changing the algorithm for passing through the transformer layers.

The main mechanism of the plugin is to skip from 65% to 70% of all transformer passes. To achieve the result, a spectral prediction method is used, which does not require additional training of the model. This makes it possible to significantly reduce the computational load while maintaining the visual characteristics of the output content.

The functionality of the extension includes support for generating images with a transparent background and the ability to use multiple reference images at the same time. The developers note that the use of the spectral prediction method gives the generation results a specific "airbrush" aesthetic, which can be useful for certain artistic tasks.

The tool is available as an open repository on GitHub. Using this method allows ComfyUI users to significantly speed up the inference process without changing the weights of the base model. This solution is aimed at those who want to optimize local workflows when working with multimodal architectures of the Qwen family.