Alibaba's Qwen team has unveiled Qwen-Image-Layered, a novel model that decomposes images into multiple RGBA (Red, Green, Blue, Alpha) layers. This layered representation allows each layer to be edited independently without affecting other content, enabling high-fidelity operations such as resizing, repositioning, and recoloring.
The model supports variable-layer decomposition, meaning it can break an image into any number of layers—for example, 3 or 8 layers—depending on the need. Additionally, decomposition can be applied recursively: any layer can be further decomposed, offering virtually infinite granularity.
Practical applications include clean object removal, distortion-free resizing, free movement of objects within the canvas, and text revision. For instance, users can recolor the first layer while keeping others untouched, or replace a subject (e.g., a girl) with another (e.g., a boy) by editing only the relevant layer.
Qwen-Image-Layered bridges the gap between raster imagery and structured, editable representations, aiming to provide intuitive, precise, and robust editing capabilities. The model is available on GitHub, Hugging Face, and ModelScope.