Qwen open-sources Qwen-Image-2.1: 7B image model with transparent layers and 10-image input
智东西 · wechat · 2026-09-21
Alibaba's Qwen team open-sourced Qwen-Image-2.1, unifying text-to-image generation and editing in a single model with a compact 7B visual generation module (32-layer Single-Stream DiT) and native transparent-image support. It scores 60.28 on Qwen-Image-Bench—first among open models, beating NanoBanana 2.0 and GPTImage 1.5 while trailing GPTImage 2 and GrokImagine 2.0.
Key capabilities
- Transparency: outputs RGBA layers, in-layer text replacement, and subject extraction for design/e-commerce workflows
- Multi-image input: up to 10 reference images for group composites, outfit try-ons, and interior design layouts
- Local editing: circle-select, brush, and independent-mask modes for precise multi-region edits
- Fidelity: facial features persist across repeated edits; product text, texture, and shape preserved in new scenes
Architecture: hybrid-granularity attention (token-level causal masks for text, chunk-level for images) plus KV Cache precomputation to cut inference memory. Also handles panoramas, infographics, and storyboard generation. Available on GitHub, ModelScope, and HuggingFace.
More from Multimodal
- P2Echo: one text-prompted model segments multi-view echocardiography with high accuracy and low cost — UBC-V · 2026-11-12
- Redditor makes fast-motion short film with H3 and ships Longshot Studio v0.18.5 with face swap — R34vspec · 2026-11-12
- Reimagining Knight Rider's opening as a Netflix drama with Claude Fable 5.2 — justin_hart · 2026-11-12
- Gemini 4 Carbon builds a full 3D aquarium in Blender and Three.js in ~6 hours — sunjiao123sun_ · 2026-11-12
- Awaken Open-Sourced: Turn a Half-Duplex Speech LLM Full-Duplex With Just a Small LoRA — alexcovo_eth · 2026-11-12
- MiniMax H3 Image targets e-commerce: sketch to 4K campaign shots in ~10 seconds — MiniMax_AI · 2026-11-12