Alibaba open-sources Qwen-Image-2.1: 7B unified image generation and editing with native RGBA
RisingSayak · x · 2026-09-21
Alibaba's Qwen team open-sourced Qwen-Image-2.1, a unified text-to-image generation and editing model with just a 7B visual generation component (32 Single-Stream DiT layers), balancing quality, inference efficiency, and versatility.
Four key improvements:
- Compact & efficient: mixed-granularity attention plus prefix KV cache reuse for strong quality at low compute
- Native transparency: generate/edit RGBA transparent images and extract subjects from photos in one model
- Versatile editing: up to 10 reference images, local edits via circles, painted annotations, or masks, with identity preservation for people and products
- Better textures: improved typography, portrait lighting, and fine details
Ships with a diffusers quick start (QwenImage21Pipeline). Independent benchmarking by RisingSayak: on a warmed A100 with two 1024×1024 references at 40 steps, DiT time dropped from 50.57s to 19.86s (2.55× speedup), at the cost of VRAM rising from 32.3G to 36.0G — with further optimization possible.
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