Qwen-Image-2.1 makes transparent PNGs natively — now non-commercial
Alibaba's 7B open image model puts out 2K RGBA cutouts and edits from 10 refs — but the weights flipped from Apache 2.0 to research-only.
Copy markdown
Alpha channels straight from the sampler
Qwen-Image-2.1 generates true RGBA through a 64-channel transparent autoencoder, so logos, icons, sprites, and product cutouts drop out of the sampler ready to composite — no rembg, no matting pass. That deletes an entire step from your asset pipeline.
2K native, one model for generate and edit
It renders 2048x2048 directly (up to 2752x1536, multiple aspect ratios), and the same 7B MMDiT edits existing images — up to 10 reference images for character/product/style consistency, mask-based local edits, and identity preservation. At 7B it runs on a single consumer GPU.
The catch: it's no longer Apache 2.0
The weights ship under the Qwen Research License — non-commercial only, with a separate license required from Alibaba for any business use. Earlier Qwen-Image releases were Apache 2.0, so anything you shipped commercially on 2.0 can't just swap this in.
Run it tonight, with a prompt-rewriter sidecar
It's live in ComfyUI and via Diffusers (QwenImage21Pipeline), vLLM-Omni, and SGLang at 40 steps (PyTorch 2.4+, Transformers 5.17+). A companion PE-I2I checkpoint (9B, Qwen3.5-VL) rewrites vague edit instructions into precise ones and keeps a source image's original language instead of defaulting to English.