Alibaba's Qwen team has launched Qwen-Image-2.1-Turbo, a new image generation and editing model based on the existing 7B visual generation architecture. The model utilizes a single-stream DiT design with 32 layers and 7B parameters, incorporating block-causal attention to enable prefix KV caching. This technical configuration allows the system to reuse text and reference-image context across denoising steps, optimizing performance with an 8-step sampling schedule.
The model is compatible with the Diffusers library, requiring specific source versions and PR #14950 for pipeline-configured sampling sigmas. It supports various generation tasks, including portraits, typography, and UI layouts, as well as editing functions like multi-reference composition. Users can process images with resolutions ranging from 2048x2048 to 2752x1536 pixels.
Alibaba Cloud Model Studio now provides access to both the Turbo and Pro versions of the model. The Turbo variant is priced at ¥0.1 per image with a request limit of 120 RPM, while the Pro version costs ¥0.25 per image with a 20 RPM limit. Consequently, the Turbo model offers a 2.5x lower cost per image and a 6x higher request rate compared to the Pro iteration.
