Alibaba's new AI splits images into Photoshop layers

Qwen-Image-Layered is a new open-source image decomposition model from Alibaba Qwen that splits any flat, single-layer image into isolated RGBA layers with transparent backgrounds.

Anyone who uses AI generators knows the frustration of editing: change one small detail and the whole scene melts, forcing you into Photoshop for tedious manual cutouts. This model tackles that by physically isolating objects and backgrounds into native layers right from the start.

What it can do

Users can explicitly control the output by specifying between 3 to 10 distinct layers through prompts or inference settings, breaking images down from broad layout zones to minute foreground details.

It also handles recursive, infinite decomposition. If a single extracted layer still contains multiple components, you can pass that specific layer back into the model to keep splitting it further into deeper sub-layers.

Once separated, elementary design operations become frictionless. You can resize an object without warping the background, drag elements freely across the canvas, erase unwanted clutter cleanly, or recolor a single object while leaving the rest untouched. Alibaba also showed it working with Qwen-Image-Edit to swap characters or modify specific text signs.

How to run it

The project is fully open-sourced under the Apache 2.0 license, with code on GitHub, paper on arXiv, and model weights hosted on Hugging Face and ModelScope.

It runs through standard Diffusers pipelines, requiring a recent Transformers release that supports Qwen2.5-VL. The model supports bfloat16 on CUDA or Apple MPS devices. While it handles both 640 and 1024 resolution buckets, Alibaba currently recommends sticking with 640 for best consistency. Interactive demos are also up on Hugging Face Spaces and ModelScope if you want to skip local setup.

Why it matters

Digital artists, graphic designers, and e-commerce creators should jump on this immediately. The biggest roadblock keeping AI art out of production workflows has always been flat, uneditable pixels that fight against professional design habits.

We think this is one of the smartest pragmatic moves in image modeling this year. Instead of racing for higher raw fidelity, Alibaba bridged the painful gap between AI generation and actual design software, turning static renders into workable assets.