It has the same architecture as Qwen-Image-2.1, with 7 billion parameters, but only eight denoising stages: the checkpoint is available on Hugging Face under a research licence.
Alibaba's Qwen team has released Qwen-Image-2.1-Turbo, an accelerated version of Qwen-Image-2.1, on Hugging Face. The model uses the same visual generation architecture of 7 billion parameters, but produces its images in just eight stages of denoising, both for creating from text and for editing existing images.
Ready to use in Diffusers
The recommended sampling schedule is embedded in the checkpoint and loaded automatically, without manual adjustment of the scheduler; the team specifies that other schedules have not been evaluated. Generation is done by default without additional guidance (CFG to 1), and a prefix caching reuses the context of text and reference images from one step to the next to save time. The model loads via a dedicated pipeline in the Diffusers library.
The sheet highlights typography and poster layout, interfaces, images with transparent backgrounds, portraits and composition from several reference images, up to four for interior design. The weights are distributed under the Qwen Research License, a research license and not an open license in the commercial sense.
Read the full article on the original source
Go to the source →