# Qwen Image Layered

> Qwen-Image-Layered is an open image model from the Qwen team, released on 19 December 2025 under Apache 2.0, that decomposes one flat RGB picture into several RGBA layers you can edit independently. Running it means Python, diffusers and a GPU. Kenerate's Image Layer tool offers layered splitting in the browser instead.

URL: https://kenerateai.com/qwen-image-layered
Publisher: Kenerate AI (https://kenerateai.com)
Last updated: 3 October 2026

## Key facts
- Released: 19 December 2025 (weights)
- License: Apache 2.0
- Size: 20B parameters (model card)
- Runs with: diffusers QwenImageLayeredPipeline, CUDA GPU
- Browser option: Kenerate Image Layer, no install
- Kenerate output: Transparent PNGs + ZIP

## Steps
1. Open Image Layer and add a picture — JPG, PNG or WebP up to 25 MB, or a sample such as the Concert poster. Sign in with a verified email to run it.
2. Choose the engine and layer count — Kenerate Layer: pick 2–8 layers (4 to start). Kenerate Layer Pro: automatic count, 1K / 1.5K / 2K output.
3. Describe the whole picture (optional) — Up to 800 characters, including parts hidden behind other things, then press 'Split into N layers'.
4. Edit and export — Use Composite, Exploded, Arrange or Grid view, move and resize layers, and download each PNG or a ZIP.

## In depth

### What is Qwen-Image-Layered?
Qwen-Image-Layered is an image model from the Qwen team that decomposes a single image into multiple RGBA layers — each with its own colour and transparency — so every layer can be edited without affecting the rest [1]. The team published the paper on arXiv on 18 December 2025 and released the weights on 19 December 2025 on Hugging Face and ModelScope, under the Apache 2.0 license [1][3]. The model card lists 20 billion parameters [2].

### How the model works
The paper, 'Towards Inherent Editability via Layer Decomposition', describes an end-to-end diffusion model that predicts several RGBA layers from one RGB input; it extends the image VAE to four channels so transparency is encoded alongside colour [3]. The number of layers is flexible — the README shows the same picture split into 3 or 8 — and any layer can be decomposed again, recursively [1].

### Running it yourself
The README's quick start installs diffusers from source plus python-pptx and psd-tools, loads QwenImageLayeredPipeline and runs it on a CUDA GPU in bfloat16, with a 'layers' setting and a resolution bucket of 640 (recommended for this version) or 1024 [1]. A Gradio app in the repository exports the layers to PPTX, ZIP and PSD files; online demos run on Hugging Face Spaces and ModelScope Studio [1].
- Python with transformers 4.51.3 or newer
- diffusers installed from GitHub
- A CUDA GPU for practical speed
- Layer count and resolution bucket set per run

### Known limits, from the authors
The README notes that the text prompt is meant to describe the whole input image — including partly hidden content — not to control what goes on each layer, and that while text-to-layers generation is supported, the released weights are tuned for image-to-layers, so text-only results are limited [1].

### Layered splitting in the browser
If you'd rather not install a 20B-parameter model, Kenerate's Image Layer tool offers layered splitting with nothing to set up: Kenerate Layer lets you pick 2–8 layers, Kenerate Layer Pro chooses the count and outputs up to 2K, and the viewer adds Composite, Exploded, Arrange and Grid views. Results download as transparent PNGs or a ZIP.

References:
[1] QwenLM — Qwen-Image-Layered on GitHub (README): https://github.com/QwenLM/Qwen-Image-Layered
[2] Hugging Face — Qwen/Qwen-Image-Layered model card: https://huggingface.co/Qwen/Qwen-Image-Layered
[3] arXiv 2512.15603 — Qwen-Image-Layered: Towards Inherent Editability via Layer Decomposition: https://arxiv.org/abs/2512.15603

## FAQ
Q: What is Qwen Image Layered?
A: Qwen-Image-Layered is an open image model from the Qwen team that splits one flat picture into several RGBA layers, each with its own transparency, so you can recolour, move, resize or delete one element without touching the rest. Its weights were released on 19 December 2025 on Hugging Face and ModelScope under Apache 2.0 (checked 3 October 2026).

Q: Is Qwen-Image-Layered open source?
A: Yes. The GitHub README states that Qwen-Image-Layered is licensed under Apache 2.0, and the weights are published on Hugging Face and ModelScope (checked 3 October 2026). Apache 2.0 is a permissive license that allows commercial use with attribution and the license notice. The paper is arXiv 2512.15603.

Q: How many layers can Qwen-Image-Layered make?
A: A variable number that you set per run. The README's example uses 4 layers and shows the same image split into 3 or 8, and any layer can be split again, recursively. Kenerate's Image Layer tool lets you choose 2–8 layers on Kenerate Layer, or have Kenerate Layer Pro pick the count, often 6–12.

Q: What do I need to run Qwen-Image-Layered locally?
A: Python with transformers 4.51.3 or newer, diffusers installed from GitHub, and in practice a CUDA GPU: the README's quick start loads QwenImageLayeredPipeline in bfloat16 on CUDA. The model card lists 20 billion parameters. If you don't have that setup, Kenerate's Image Layer runs layered splitting in a browser instead.

Q: Can I export the layers as a PSD?
A: From the official Qwen-Image-Layered demo app, yes: its README says it exports layers to PPTX, ZIP and PSD files. Kenerate's Image Layer doesn't export PSD; it gives you each layer as a transparent PNG, a ZIP of all of them, and a PNG of any new composition you make in Arrange mode.

Q: Does the prompt control what goes on each layer?
A: No. The Qwen-Image-Layered README says the text prompt should describe the whole input image, including partly hidden content, and isn't designed to assign content to specific layers. Kenerate's optional 'Describe the picture' field, up to 800 characters, works the same way: describe the scene and let the engine decide the layers.

Q: Can I try layered splitting online without installing anything?
A: Yes. Qwen hosts demos on Hugging Face Spaces and ModelScope Studio, and Kenerate's Image Layer tool runs in any browser: upload a picture, choose 2–8 layers or Kenerate Layer Pro's automatic count, and download transparent PNGs or a ZIP. New accounts get starter credits, and there is no subscription required.

## Related pages
- Separate Image Into Layers: https://kenerateai.com/separate-image-into-layers
- AI Background Remover: https://kenerateai.com/free-ai-background-remover
- Transparent Background Maker: https://kenerateai.com/transparent-background-maker
- AI Image Editor: https://kenerateai.com/ai-image-editor
- AI Image Upscaler: https://kenerateai.com/ai-image-upscaler

## Sources
- QwenLM — Qwen-Image-Layered on GitHub (README): https://github.com/QwenLM/Qwen-Image-Layered
- Hugging Face — Qwen/Qwen-Image-Layered model card: https://huggingface.co/Qwen/Qwen-Image-Layered
- arXiv 2512.15603 — Qwen-Image-Layered: Towards Inherent Editability via Layer Decomposition: https://arxiv.org/abs/2512.15603
- Kenerate Image Layer: https://kenerateai.com/app/tool/image-layer