Qianwen said it open-sourced Qwen-Image-2.1 today. The 7B-parameter image model
combines text-to-image generation and image editing, is optimized for inference
efficiency and low cost, and natively supports generation and editing of
transparent images. Main features include a compact, inference-optimized
architecture that balances quality and compute cost; native transparent-image
generation with transparent-layer editing and matting; multi-reference editing
support (up to 10 reference images) to boost local edits, portrait and product
fidelity across multiple edit tasks; and improved text layout, lighting and
fine-detail rendering.