Alibaba's Qwen team has released Qwen-Image-2.1, a unified text-to-image generation and image editing model that consolidates previous separate checkpoints into a single 7B parameter diffusion transformer.
- The model uses a 7B DiT with block-causal attention and an 8B Qwen3-VL encoder to process text and condition images.
- It supports native RGBA transparency, multi-reference editing with up to 10 images, and local region edits while preserving identity.
- A mixed-granularity attention mechanism allows prefix KV cache reuse, computing text and reference inputs only once per generation step.
- On Qwen's in-house Qwen-Image-Bench, the model scores 60.28, outperforming listed open-weight models like FLUX 2 Max (55.33) but trailing closed models like GPT Image 2.5 Sunburst (67.01).
- The release includes Day 0 support for Diffusers, ComfyUI, vLLM-Omni, and SGLang, with commercial use requiring a separate license.
The unified architecture aims to provide a cost-effective solution for both generation and editing tasks while maintaining high aesthetic quality and transparency support.