Qwen: qwen-image-2.1-turbo

Qwen Image 2.1 Turbo is a distilled accelerated checkpoint in the Qwen-Image-2.1 series, built on the same 7B vision generation architecture as the base model. It uses knowledge distillation to compress denoising to an 8-step sampling schedule at CFG 1.0 while keeping 2K-level image generation. The model runs at BF16 native precision and is about 14.2GB in size. Weights are available on Hugging Face and ModelScope and can be loaded through the Diffusers QwenImage21Pipeline. It supports text-to-image generation and image editing driven by natural language instructions, with up to 4 additional reference images per editing request. It also supports RGBA output for transparent-background images. For editing, it offers stronger restoration of facial details and better consistency of product text and textures across new scenes. The recommended generation resolution is about 2 megapixels, such as 2048x2048 at 1:1. This model is best suited to high-frequency local image generation workflows and rapid iteration, including product visual design for posters, concept art, and ad assets, as well as e-commerce product imagery and model outfit changes. It must use its dedicated 8-step sampling schedule, as the standard default schedule produces different images. At 4-bit quantization and below, output deviates noticeably from the original, and higher-quality quantized text encoders are recommended because heavily quantized encoders degrade output. It is released under the Qwen Research license, which carries restrictions on commercial use.