Flat images are easy to make and painful to reuse. Change one element — swap a product, recolor a shape, move a character — and you're either hunting for the original PSD or masking pixels by hand.
Layer decomposition attacks the problem at the source: instead of treating an image as a grid of pixels, an AI model splits it into semantic, editable pieces.
What the model does
Qwen-Image-Layered (released by Alibaba's Qwen team in December 2025, arXiv 2512.12299) decomposes a single image into multiple RGBA layers. Each layer is a full-resolution, transparent-background component — object, text, background — that you can move, recolor, or replace independently, then recompose back into the original image.
The interesting part is the training objective: the model learns to produce layers whose recomposition reproduces the input, so decomposition quality is directly measurable (reconstruction fidelity + semantic separation).
What that unlocks in practice
- Design reuse: pull a product shot out of a banner without a clipping path
- Transparent PNG export: every layer exports with alpha, ready for Photoshop, Figma, or Canva
- Batch asset work: e-commerce and ad teams can regenerate variations from one master image
- Downstream editing pipelines: layers are structured data, so they feed cleanly into editors and automation
Try it
We run a hosted implementation at ImageLayered: upload a flat image, get named, editable layers back. Quick mode starts at \$0.05 per layer, Precision mode decomposes up to 16 layers, and new accounts get free credits to test it. If you want to go deeper, the core model (Qwen-Image-Layered) is available for self-hosting.
If you build anything with image layering — or you have opinions on where decomposition beats segmentation — I'd love to hear about it in the comments.
Top comments (0)