What is Qwen Image Layered?
Qwen Image Layered is a diffusion model built on Qwen-Image. You give it a picture, the number of layers you want and ideally a caption of the whole scene, and it returns RGBA layers that stack back into the original. Any layer can be split again, and the official app saves the result as PPTX, ZIP or PSD.
It is the only version the Qwen team has released, and the weights have not changed since launch. Unlike prompt editors such as Nano Banana 2, it separates an image rather than changing it on request, but it redraws every layer, so fine detail and text can come out softer than the original.
Who is Qwen Image Layered best for?
Qwen Image Layered is best for designers and developers who need to turn a flat image into separate editable layers, so one element can be recoloured, moved or removed without touching the rest.
Editing one element of a flat image
Each object lands on its own RGBA layer, so recolouring or replacing it leaves the rest of the picture untouched.
Moving and resizing objects
Layers can be moved around the canvas or resized without distortion, and unwanted objects deleted cleanly.
Getting images into slides or Photoshop
The app in the official repo exports the layers as PPTX, ZIP or PSD files for editing in familiar tools.
Products that need layer splitting
Apache 2.0 weights that run in diffusers or vLLM-Omni can sit inside your own pipeline with no fees to the makers.
What's included with Qwen Image Layered?
Qwen Image Layered splits an image into as many RGBA layers as you choose, splits any layer again, and exports to PPTX, ZIP or PSD.
- Variable layer count
- You choose how many layers to split an image into; the official app allows 2 to 10, with 4 as the default.
- Recursive splitting
- Any layer can be decomposed again, so a crowded foreground can be broken down further.
- Export to PPTX, ZIP or PSD
- The local app in the GitHub repo saves layers in all three formats; the Hugging Face demo saves PPTX and ZIP.
- Layer editing tools
- Scripts in the repo edit a single RGBA layer with Qwen-Image-Edit and combine layers back into one image.
- Local and pipeline use
- Runs through diffusers on an NVIDIA GPU or through vLLM-Omni, and ComfyUI ships a native workflow with a lower-memory fp8 file.
- Free online demos
- Hugging Face Spaces and ModelScope Studio host the model, so you can test it without a GPU.
How does Qwen Image Layered work?
Add your image
Upload a picture to the demo or pass it to the diffusers pipeline; 640-pixel input suits this version best.
Set layers and a caption
Pick how many layers you want and describe the whole image, including anything hidden behind other objects.
Edit and export
Download the layers as PPTX or ZIP (PSD in the local app), or split any single layer again.
How much does Qwen Image Layered cost?
Qwen Image Layered is free: the Qwen team released its weights under the Apache 2.0 license, so you pay only for your own GPU, and free online demos run on Hugging Face Spaces and ModelScope Studio.
- 20B-parameter model on Hugging Face and ModelScope
- Runs with diffusers on an NVIDIA GPU, or with vLLM-Omni
- Local app exports layers as PPTX, ZIP or PSD
- Upload an image and get its layers back
- No install or GPU of your own
- Runs on Hugging Face's shared ZeroGPU hardware
There is no paid plan or hosted API from the maker: Alibaba Cloud's Model Studio price list does not include the model, and Hugging Face lists no inference provider for it. The Hugging Face demo shares a daily ZeroGPU quota: 2 minutes of GPU time without an account, 5 minutes on a free account and 40 minutes on Hugging Face PRO. The online demo also runs on ModelScope Studio. Full pricing on huggingface.co. We checked these prices on .
How do readers rate Qwen Image Layered?
No one has rated Qwen Image Layered yet. Be the first.
Sign in to rate. One rating per account.
Qwen Image Layered pros and cons
Qwen Image Layered is free, open and truly layer-aware, but users find its layers soft and slow to make, so it suits rough splits more than finished work.
Pros
- Free under Apache 2.0, and one of the few open models from a major lab that returns real transparent layers, a point both Hacker News and Reddit users made.
- Each layer is a separate RGBA image, so an edit to one element cannot spill into the rest, and stacking the layers rebuilds the picture.
- Users who tested it get the cleanest splits on poster-style images with simple shapes and strong contrast.
- On the Qwen team's Crello layer test it beat LayerD, the earlier open method, with 0.87 alpha soft IoU against 0.75.
Cons
- Output quality is the main complaint: Reddit users report soft, blurry or grainy layers, lost detail and garbled text, because the model redraws each layer instead of cutting out the original pixels.
- Slow and heavy: the 20B model needs a large GPU, users report minutes per image, and generation slows as you add layers.
- It sometimes groups things oddly, splitting by object rather than by depth and pushing images towards a poster-style layout.
- Only one version exists: the weights have not changed since release, this version is tuned for 640-pixel input, and making layered images from a text prompt alone works poorly.
Is Qwen Image Layered open source?
Yes. Qwen Image Layered's code is public under the Apache 2.0 licence, so you can run it on your own server.
On GitHub
2,185stars
- Last commit
- 31 Dec 2025
- Repository age
- 11 months
- Licence
- Apache 2.0
- Self-hosted
- Yes
- Repository
- QwenLM/Qwen-Image-Layered
What are the best Qwen Image Layered alternatives?
Qwen Image Layered is an open source alternative to Seedream 5.0 Pro.
Questions about Qwen Image Layered
What is Qwen Image Layered?
Qwen Image Layered is the Qwen team's open 20B-parameter model that splits a flat picture into separate transparent layers, so designers and developers can edit one element without touching the rest.
Who is Qwen Image Layered for?
Qwen Image Layered is best for designers and developers who need to turn a flat image into separate editable layers, so one element can be recoloured, moved or removed without touching the rest.
Is Qwen Image Layered free?
Yes. Qwen Image Layered has a free Open weights plan: 20B-parameter model on Hugging Face and ModelScope.
Is Qwen Image Layered open source?
Yes. Qwen Image Layered's code is public under the Apache 2.0 licence, so you can run it on your own server.
What are the best Qwen Image Layered alternatives?
The closest alternatives on Recaply are Seedream 5.0 Pro and Reve.
Where can I try Qwen Image Layered?
On its official website, github.com.
Can I use Qwen Image Layered commercially?
Yes. Qwen Image Layered's code and weights are released under the Apache 2.0 license, which allows commercial use, changes and redistribution as long as you keep the license and its notices with the code.
Can Qwen Image Layered make layered images from a text prompt?
Only weakly. The released weights were tuned to split an existing image into RGBA layers, and the Qwen team says text-to-layers results are limited. Your prompt should describe the whole input image, including parts hidden behind other objects, rather than each layer.
Why do my Qwen Image Layered layers look blurry?
Mostly resolution and redrawing: this version is tuned for 640-pixel input, and the model regenerates every layer rather than copying pixels. ComfyUI suggests 1024-pixel input for higher-resolution output, and a caption describing the whole image helps the split.
Can I run Qwen Image Layered in ComfyUI?
Yes. ComfyUI has a native image-to-layers workflow. The default bf16 file needs a lot of GPU memory, an fp8 file cuts that, and ComfyUI warns the model is slow at its original 50 steps.
Is there a newer version of Qwen Image Layered?
No. The Qwen team has released one version, its weights have not changed since launch, and it is still the only layer model on the team's Hugging Face page, so the first release is the one you get.
Our verdict
Qwen Image Layered is worth a try if you need layers from a flat image and can live with rough output.
It is free under Apache 2.0 and one of the few open models that returns real RGBA layers, but users find the layers soft, slow to make and oddly grouped outside poster-style images, and no improved version has followed.
- Skip it for
- Anyone who needs print-ready layers or sharp text, or who has no large GPU and needs more than the free demo's daily quota.
- Check first
- Run a few of your own images through the free Hugging Face demo before you build it into a pipeline.
Qwen-Image-3.0
Bonsai Image 4B
FLUX 3 Image Klein
Hunyuan Image 3.0
ChatGPT Images 2.5
GLM-Image