Alibaba Open-Sources Qwen-Image-2.1 with Native Transparency and Multi-Image Reference Support
Alibaba's Qwen team has open-sourced Qwen-Image-2.1, a 7B-parameter unified model for image generation and editing. The release features native transparent layer generation (RGBA), refined multi-lingual text rendering, and support for up to 10 reference images during image-to-image workflows. By combining text-to-image creation, localized editing, and native transparency into a compact 7B model, Qwen-Image-2.1 streamlines professional visual workflows for local deployment. It significantly lowers compute costs while enabling creators to generate graphic assets ready for composition without external background-removal tools. The architecture leverages mixed-granularity attention and key-value (KV) cache reuse to treat reference images as static context, optimizing memory usage and inference speed. However, community members noted that Qwen-Image-2.1 is governed by a custom, more restrictive license than the Apache 2.0 license used in earlier Qwen releases.
## BACKGROUND
Standard text-to-image diffusion models generate flat RGB images, requiring separate computer vision algorithms (such as matting or background removal) to create transparent PNGs. Native transparency models integrate alpha channel predictions directly into the diffusion process, allowing the neural network to output isolated foreground elements or edit specific image layers with greater precision.