One model now covers the cutout pass.
Qwen released Qwen-Image-2.1 on September 20. Its visual generation component has 7 billion parameters and 32 single-stream diffusion transformer layers. The release combines text-to-image generation, image editing, subject extraction, and transparent image work in one model.[1][2]
The useful change is the output contract. A prompt can ask for a regular image or an image with an alpha channel. The same model can edit a transparent layer or extract a subject from a photograph. Qwen also says the model accepts up to 10 reference images for composition and editing.[1]
These are vendor claims and examples. They do not prove that every subject, edge, type treatment, or reference set will survive your workflow.
RGBA removes a step. It does not remove inspection.
A white background and transparent pixels can look identical on a white page. The difference appears when the asset moves onto black, color, video, print, or a compositing tool. Pale halos expose a bad matte. Missing semi-transparent pixels turn hair, smoke, glass, and soft shadows into hard stickers.
If the edge has not seen both black and white, it has not passed the booth.
Check the file itself. Confirm that the output has an alpha channel rather than a baked checkerboard. Inspect the silhouette on light and dark backgrounds. Verify width, height, crop, and empty padding. Then keep the model name, workflow, prompt, input references, seed, and license with the asset.
Day-zero support still needs a local preflight.
The Qwen repository lists day-zero support in Diffusers and ComfyUI. The Comfy Org model card provides separate diffusion model, text encoder, and VAE files. It also links text-to-image and image-edit workflow templates.[2][3]
The model card labels the release with the Qwen Research license. Read that license before client or product use. The card also says no hosted inference provider currently deploys its repackaged model, so that page is a download and workflow route rather than a push-button hosted endpoint.[3]
Ship the asset with four checks.
- Alpha: inspect the file mode or channel list. A checkerboard drawn into RGB pixels does not count.
- Edge: view the subject on white, black, and one loud color. Look for fringe, holes, clipped detail, and dirty semi-transparent pixels.
- Canvas: confirm dimensions, crop, padding, orientation, and the intended anchor point.
- Receipt: record the model, weights, workflow revision, prompt, references, seed, license, and human review result.
Native alpha can make asset work faster. The saved time belongs in composition and review, not in pretending the mask cannot fail.