Qwen-Image-Layered: AI Can Now Turn Any Photo into an “Almost PSD” by Decomposing Images into RGBA Layers
The Qwen team has released Qwen-Image-Layered, a model that does what designers have been dreaming about for years: it takes an ordinary flat image and decomposes it into multiple transparent RGBA layers, allowing each element to be moved, recolored, removed, or scaled independently without breaking the rest of the image. The repository explicitly emphasizes “inherent editability”: editing happens “physically” at the layer level, rather than through careful attempts to repaint pixels on a single canvas.
The key insight lies in how the model learns this behavior in the first place. In their paper, the authors describe a pipeline that extracts and annotates multilayer data from Photoshop PSD files, because high-quality “layered” datasets are extremely scarce. This explains the result: the model has learned to infer the typical structure of a scene — where the background is, where the main object is, where text or decorative elements are, and where overlaps occur — and outputs not just masks, but full layers with alpha channels.
In practice, this is highly applicable: there is a demo available on Hugging Face Spaces, and the project’s tools support exporting the result to PSD (as well as to other convenient formats such as PPTX or zip), so the output can be further edited manually using familiar software.
