Alibaba introduced the multimodal model Qwen-Image-2.1

Company Alibaba released Qwen-Image-2.1 — a compact neural network with 7 billion parameters, designed for image generation and editing. The model combines the functions of creating visual content from text descriptions and tools for its subsequent correction within a single architecture. Key feature of the system is the support of native transparency, which allows generating objects with an alpha channel without the need for additional post-processing. The model also supports the use of up to 10 reference images simultaneously, ensuring high accuracy when editing specific elements of the composition or preserving style. Developers optimized the architecture for working with visual data, while maintaining a compact size characteristic of the Qwen series. This allows the model to be launched on consumer equipment with lower requirements for video memory compared to large diffusion models, while maintaining flexibility in generation management. Integration of editing and generation functions into one model simplifies pipelines for designers and developers working with AI graphics. Detailed technical specifications and examples of work are available in the official blog of the project, where approaches to training the model are also discussed. Release of Qwen-Image-2.1 continues Alibaba's strategy of developing effective open models that can compete with proprietary solutions in computer vision tasks. Discussion of technical aspects of the release is actively being conducted by the community on the Hacker News platform.