Qwen releases model that splits images into layers
Alibaba's Qwen team has introduced Qwen-Image-Layered, which breaks a single image into separate RGBA layers that can each be edited on their own, so changes to one part leave the rest untouched.
The Qwen team announced Qwen-Image-Layered, a model that takes a flat raster image and decomposes it into multiple RGBA layers. Each layer holds a distinct semantic or structural component. According to the team, isolating components this way gives the image "inherent editability": an edit applied to one layer cannot spill over into the others.
The blog post shows several examples. Users can recolor one layer, swap a girl for a boy in another, or change embedded text to "Qwen-Image," and the rest of the image stays as it was. The team also says the layered format naturally supports basic operations such as cleanly deleting objects, resizing them without distortion and moving them anywhere on the canvas.
The number of layers is not fixed. The post shows the same image split into either 3 or 8 layers, and any layer can be decomposed again. The team describes this recursive process as enabling "infinite decomposition."
The authors present the model as a bridge between raster imagery and structured, editable representations. A technical paper, "Qwen-Image-Layered: Towards Inherent Editability via Layer Decomposition," is listed on arXiv (2512.15603). Its authors include Shengming Yin, Chenfei Wu, Junyang Lin, Jingren Zhou and Heung-Yeung Shum.
The announcement does not give quantitative benchmarks, model size, licensing terms or details on how to access the model.