Qwen-Image-Layered - Turn flat images into multi-layer editable assets

Hi everyone!

This is a really interesting release from the Qwen team. The model natively decomposes images into layers, which is basically the foundation for any precise editing work.

It is fine-tuned based on Qwen-Image. Right now, the Hugging Face demo allows splitting up to 10 layers, which is usually enough for most workflows.

The model size is a bit heavy. Hopefully, we will see a lighter or distilled version in the future!

Wow, this is absolutely huge!

Can't wait for SaaS video-editors to appear online to take advantage of this opportunity! 🚀

Hey, so cool to see my ex-company’s product on here! Qwen-Image-Layered is such a cool idea — as someone who’s suffered through way too many manual cutouts, turning images directly into editable layers feels like magic. Really smart approach.

Major props for focusing on such a specific yet painful problem. It's refreshing to see a bigger team move this nimbly and deliver something so genuinely useful.

One tiny thought for the future — if the AI could auto-tag layers while decomposing (like "person-foreground", "text-header", "sky-background"), it’d be even easier to navigate, especially when there are lots of layers.

Anyway, super stoked to see this launch. Congrats to the team! Looking forward to catching up 👏

This was so needed! Using LLM image generators was really limiting without layers

This is a solid step forward for practical image editing. Native layer decomposition is exactly what’s been missing for precise control, not just “generate and hope.”

Agreed on the model size — a distilled version would make this much easier to adopt in real-world workflows.

Oh man...I can't wait to use this. I have been using inkscape to do all this manually for so long now. I use it to make some really cool 3D printed art from comics and stuff. I can't wait to try this! Like, I am genuinely super excited! If I was a dog, my tail would be straight up vibrating at this idea! I hope my friends and family don't get mad at all the 3D printed art I start churning out. I hope to someday start making 'limited edition' comic book covers for people. I think this will get me really close.

@junyang_lin I wanted to try, but don't you release the windows version?

It is a significant milestone because it addresses the "tangled pixel" problem that has plagued AI image editing. Traditionally, AI edits images by regenerating the entire canvas (or a masked area), which often leads to "drift"—where the surrounding parts of the image change unintentionally.

By shifting from flat pixels to RGBA layers, Qwen-Image-Layered brings professional Photoshop-style logic to generative AI.

Can't wait to see APIs built on top of it

Interesting! Soon China will overtake the US in the number of AI projects)