Qwen-Image-Layered image model

Qwen-Image-Layered, released by Alibaba in December 2025, is the image layers editing version of Qwen-Image model. It breaks a flat image into layers you can edit on their own. Each part, like objects, backgrounds or effects, goes into its own RGBA layer with transparency. So when you edit one part it won’t mess with the rest. Kinda like what you'd do in Photoshop.

Instead of changing pixels straight on the image, the model gives you a clean stack of layers. You can move, recolor, resize or delete stuff without messing up nearby things.

The model's built on a paper from Dec 2025. It uses a diffusion method and a RGBA-VAE to link normal images with layered ones. It also has a setup that lets you pick how many layers to use. They trained it in steps, starting from a model that already knew how to handle images.

This setup gives cleaner edits than regular pixel editing. It’s handy when dealing with complicated scenes.

You start by picking an RGBA image. The model breaks it into layers based on what’s in it. You can tweak any single layer - change colors, swap objects, delete stuff - while leaving the rest untouched.

You can pick how many layers you want, like 3, 4, or 8. And if needed, break down one layer into even smaller ones. That gives you more control over edits and layout.

Model’s up on Hugging Face under Apache-2.0 license. There are GGUF versions out on day 0, and 13gb for Q4_K_M isn't small but should work ok on a 16GB card.