Article
Qwen-Image-2.1-Turbo Open-sourced, image generation reduced from 40 steps to 8 steps
Beating AI breaking news: Alibaba Tongyi Qianwen has released Qwen-Image-2.1-Turbo and opened its model weights. This is an accelerated version of Qwen-Image-2.1, retaining the 7000000000-parameter image-generation architecture. Official examples for the original version use 40 denoising steps, while Turbo reduces this to 8 steps, a reduction of 80%. Denoising is the process by which a model progressively generates an image from noise. However, a reduction of 80%in steps does not mean actual processing time is also reduced by 80%.
The new model can still generate 2K images and supports text-guided image editing, multi-image references, and transparent-background output. Developers can run it locally through Hugging Face's Diffusers library, which automatically uses the recommended 8-step sampling configuration. The API for Turbo and Pro are also now available. At Alibaba Cloud Bailian's list price for the Beijing region, Turbo costs 0.1 yuan per image, while the Pro version costs 0.25 yuan, making the former 60%cheaper.
Some users have reported noticeably faster generation, but assessments of image quality vary. After testing it on RTX 5090, one user found that Turbo excessively sharpened skin texture when generating people. There are currently insufficient independent evaluations to demonstrate that its image quality matches the original version.
Notably, Turbo uses the Qwen Research License. Although the model weights can be downloaded, free use is limited to non-commercial research and evaluation. Commercial development using the weights still requires separate authorization.
Original link https://m.theblockbeats.info/flash/371104