Qwen-Image-2.1 uses one 7B visual generator for text-to-image generation and image editing. It can generate and edit transparent RGBA assets, accept up to ten reference images, and target local edits with circles, painted annotations or masks. Diffusers, ComfyUI, vLLM-Omni, SGLang and LightX2V added day-zero support. The weights are public, but the Qwen Research License limits them to non-commercial research and evaluation.
Qwen releases Qwen-Image-2.1 for unified generation and editing
2026-09-20Verified announcementTRIAL
Qwen-Image-2.1 uses one 7B visual generator for text-to-image generation and image editing. It can generate and edit transparent RGBA assets, accept up to ten reference images, and target local edits with circles, painted annotations or masks. Diffusers, ComfyUI, vLLM-Omni, SGLang and LightX2V added day-zero support. The weights are public, but the Qwen Research License limits them to non-commercial research and evaluation.
Imported from the existing research archive. Migration preserves the recorded judgment and is not a fresh verification.
What this means for you
For teams building local creative-asset pipelines, one checkpoint can now cover generation, composition, editing and transparent output without switching models. Run a non-commercial proof of concept for typography, identity drift, alpha-edge quality, repeated-edit degradation and target-hardware latency. Commercial teams should treat the model as WATCH until they obtain separate license terms.
Evidence & related material
Technical details & scores
| Architecture | 7B visual generator with 32 single-stream DiT layers, mixed-granularity attention and reusable prefix KV state |
|---|---|
| Capabilities | unified text-to-image and editing; native RGBA generation/editing; up to 10 reference images; circles, painted annotations and masks for local edits |
| Resolution | native 2K; official examples include 2048x2048 and multiple landscape/portrait aspect ratios |
| Runtime Support | day-zero Diffusers, ComfyUI, vLLM-Omni, SGLang and LightX2V integrations |
| Serving Measurement | SGLang reports 8.23s generation and 9.85s editing medians at 1024x1024/40 steps on one RTX PRO 6000 Blackwell, with 40.1 GiB peak device memory; DiT offload lowers peak to 26.1 GiB with higher latency |
| License | Qwen Research License; non-commercial research and evaluation only; commercial use requires a separate license |
| Benchmark Status | visual-quality and identity-preservation claims are vendor-reported; no independent quality evaluation was available at scan cutoff |
Technical impact 4
Engineering value 4
Adoption signal 3
Maturity 3
Verification cost 3
Risk 4