Alibaba unveils enhanced image generation model with faster processing

Alibaba has launched Qwen-Image-2.1-Turbo, an optimized version of its Qwen-Image-2.1 model, designed to generate and edit images in just 8 denoising steps compared to the standard 40-step process. This improvement allows developers to achieve results with 5 times fewer steps using the same 7B architecture, along with a hosted API option. The model retains the 7B visual generation architecture and is compatible with the Diffusers framework. It features a recommended 8-step sampling schedule and uses CFG=1 by default. The Turbo variant supports image editing with features like single-image transformation and multi-reference composition, while the base model allows up to 10 reference images. Alibaba Cloud Model Studio now hosts both Turbo and Pro versions, with Turbo offering a lower cost of CNY 0.1 per image and a higher request rate of 120 RPM compared to Pro’s CNY 0.25 per image and 20 RPM.