Qwen-Image-2.1

Image generation and editing in one 7B model. Qwen-Image-2.1 is a 32-layer single-stream DiT (7B parameters in the visual generation component) with a Qwen3-VL 8B text encoder and a 64-channel RGBA VAE at 16x spatial compression. It generates and edits transparent RGBA images natively (the Layered lineage folded in), takes up to 10 reference images per edit, supports local edits via circles, painted marks, or separate masks with identity preservation for people and products, and outputs up to 2752x1536 natively (default 2048x2048, 40 inference steps). Efficiency comes from mixed-granularity attention (token-level causal masking for text, chunk-level for image) plus prefix KV cache reuse across denoising steps.

License is the catch. Qwen-Image 1 shipped under Apache 2.0; this release drops to the Qwen Research License: non-commercial use only, and any commercial purpose requires a separate license from Qwen ([email protected]). The site open-weights classification counts it as open, but research and evaluation are the only permitted uses out of the box. Budget for the commercial license before building a product on it.

Ecosystem, day one. diffusers QwenImage21Pipeline, native ComfyUI with official workflows, vLLM-Omni (FP8, CUDA graphs), SGLang, LightX2V, and stable-diffusion.cpp; AMD ROCm and FlagOS multi-chip support are noted in the README.

What to treat as claims. The Qwen-Image-Bench score (60.28, rank 6, behind GPT Image 2.5 at 67.01 and ahead of Nano Banana 2 at 59.83) is Qwen own benchmark - self-scored, no independent eval yet. The memory figure is community-reported, not official: a Q8 runtime at 15.6 GB VRAM total (7.3 DiT + 7.7 text encoder + 0.6 VAE) on RTX 3090-class cards, with diffusers enable_model_cpu_offload() path for smaller ones. Community reports of a periodic dot pattern in midtones exist but are unverified; treat content marking as unknown until confirmed.