LayerGrab

What Is Qwen-Image-Layered?

Qwen-Image-Layered is an open-weight image model from the Qwen team that decomposes one image into several RGBA layers, so each part can be moved, resized or recoloured on its own. Its weights were released on December 19, 2025 under the Apache 2.0 licence, and it runs on your own GPU.

Disclosure: this page is published by LayerGrab, a hosted tool that also splits images into layers. Facts about the model come from its own documentation, linked at the bottom and checked on 2026-09-29.

Last updated 2026-09-29

Key facts

What it does
Splits one image into several RGBA layers. The same step can be applied again to any layer.
Licence
Apache 2.0
Released
Paper December 18, 2025; weights December 19, 2025; online demo December 22, 2025
Size
20B parameters, BF16 weights
How you run it
Python with the diffusers pipeline QwenImageLayeredPipeline on a CUDA GPU (transformers 4.51.3 or newer)
Number of layers
You choose it with the layers setting (the examples use 3, 4 and 8)
Resolution
Buckets of 640 or 1024; the README recommends 640 for this version
Text prompt
Describes the overall content of the image. The README says it is not designed to control the semantic content of individual layers.
Export
RGBA images; the README shows exporting the layers to PPTX

What the prompt does, and does not do

It is easy to assume you can type “separate the logo” and get the logo. The README says otherwise: the prompt describes the whole image, including parts that may be hidden, and is not meant to control which content ends up on which layer.

What you control is the number of layers. If you need one particular element out of a picture, a tool that takes an instruction, such as separating one named object, fits better.

Compared with LayerGrab

Qwen-Image-LayeredLayerGrab
CostFree weights; you pay for your own GPU timeCredits per layer returned; samples free without an account
SetupPython, a CUDA GPU with enough memory for a 20B modelBrowser, or the Photoshop, Figma, Sketch and Chrome plugins
How many layersYou set the number before runningThe model decides from the picture, up to 17 including the background
Choosing what to separateNot what the prompt is for (see README)Name it, e.g. "only the price tag", and get just that
Layer names and positionsNot documentedEach layer named in English, with its position in the image
OutputRGBA images, PPTX exportPNG, ZIP, layered PSD, or layers placed straight into your design
Where it runsOn your machine; images never leave itOn LayerGrab servers; results kept 30 days

Choose the open model when

You have a suitable GPU, want images to stay on your own machine, are building layer decomposition into your own pipeline, or are doing research. It costs nothing per image once it runs.

Choose LayerGrab when

You want layers without setting up a model, need to name the one element you want, or want the layers placed in Photoshop, Figma or Sketch. It is not the right choice if images must never leave your machine.

Try LayerGrab on a sample

Frequently asked questions

Is the model free to use?

The weights are free to download under Apache 2.0, which allows commercial use. Running the model is not free in practice: it has 20B parameters and needs a CUDA GPU, so you pay for the hardware or cloud GPU time.

Can I tell the model which object to separate?

Not directly. Its README says the text prompt describes the overall content of the image and is not designed to control what goes into each layer. You choose how many layers to produce instead.

Does LayerGrab use Qwen-Image-Layered?

No. LayerGrab runs a different hosted model. This page compares the two so you can pick the one that suits how you work.

Can I try it without installing anything?

Yes. The README links an online demo on Hugging Face Spaces and ModelScope Studio, available since December 22, 2025.

Sources