Turning flat art into editable components
Creates a practical starting stack for repositioning, recoloring or removing major visual elements.
Independent tool overview
Qwen Image Layered at a glance
Qwen-Image-Layered is a 20-billion-parameter open image model from the Qwen team that predicts multiple editable RGBA layers from a single flat RGB image.
Visit the official Qwen Image Layered site ↗
- Product type
- Open image-to-layer decomposition model
- Developer
- Qwen team
- Model size
- 20B parameters, BF16
- Output
- Variable multi-layer RGBA images
- License
- Apache 2.0
- Released
- December 2025
- Pricing checked
- August 31, 2026
Overview
What Qwen Image Layered is
Qwen-Image-Layered addresses a specific design problem: raster images normally fuse every object, texture and background into one canvas. The model decomposes an input image into a variable number of transparent RGBA layers so users can move, resize, recolor, remove or separately edit components.
The released repository includes model weights, Diffusers inference code and a Gradio interface that can export layer results as PSD, PPTX or ZIP files. It also supports recursive decomposition, where a predicted layer can be run through the model again for finer separation.
This is a developer-oriented model, not automatic recovery of an image's original design file. Layer boundaries, ordering and content are model predictions; pixels hidden behind foreground objects may be synthesized, and complex transparency, shadows, reflections, text and fine edges can require manual repair.
Use cases
Who Qwen Image Layered is best for
The strongest fit depends on the job you need the product to complete, not the size of its feature list.
Design-tool and research prototypes
Useful to developers testing layer-aware editing interfaces, presentation exports or creative pipelines.
Controlled image revisions
Separating a target element can reduce unintended changes to the rest of a composition during later edits.
Capabilities
Core Qwen Image Layered features
Image-to-RGBA decomposition
Predicts multiple transparent layers whose composite approximates the original RGB image.
Variable layer count
Supports different requested decomposition depths rather than a single fixed output count.
Recursive decomposition
A selected layer can be decomposed again to create a finer hierarchy.
Independent layer operations
Predicted elements can be moved, resized, hidden or recolored without directly changing other layer files.
PSD, PPTX and ZIP export
The repository's Gradio application can package output for downstream editing and presentation workflows.
Text-conditioned inference
A prompt can describe the overall image and partially occluded content, but is not intended to specify each layer's semantics directly.
Open implementation
Weights, example code, paper and Apache-2.0 repository are publicly available.
Process
How the Qwen Image Layered workflow works
Step 1
Confirm rights and choose an input
Use an image you may process and modify; avoid confidential or personal imagery in third-party demos.
Step 2
Prepare the runtime
Install the documented Transformers and Diffusers versions plus the export dependencies, then load the BF16 model on suitable hardware.
Step 3
Set the decomposition
Choose layer count, resolution and a factual overall-image prompt, including important content that may be partly hidden.
Step 4
Inspect every output layer
Check edges, transparency, ordering, duplicated details, synthesized occlusions, text and whether the recomposite matches the source.
Step 5
Repair and export
Edit or recursively decompose selected layers, recompose for QA, then export PSD, PPTX or ZIP while retaining the original image.
Cost
Qwen Image Layered pricing and free plan
The official code and weights are available under Apache 2.0 with no model-license fee listed. Real cost comes from GPU hardware or hosted compute, storage, engineering and any third-party demo or deployment service; Qwen does not publish a standalone paid plan for this repository.
Model and code
Free under Apache 2.0
Download and run the official repository subject to the license.
- Public weights and source
- No official per-image price listed
- Users supply compatible hardware and dependencies
- Review license and notices before redistribution
Public demo
Free, limited availability
Try the linked Hugging Face Space or ModelScope Studio when capacity is available.
- Queues and resource limits may apply
- Do not upload sensitive images
- A demo is not a production SLA
Self-hosted deployment
Compute-dependent
Run locally or on cloud GPU infrastructure.
- 20B BF16 model
- GPU memory and inference time depend on configuration
- Add storage, monitoring, security and engineering costs
Pricing checked . Check current pricing at the source ↗
Assessment
Qwen Image Layered strengths and limitations
Where it stands out
- Solves a concrete editability problem that ordinary flat-image generators do not.
- Produces real transparent layer files suitable for normal downstream operations.
- Variable and recursive decomposition can adapt output granularity.
- Repository includes working inference, export and editing utilities.
- Apache 2.0 licensing is permissive for research and product experimentation.
- The paper and implementation make the technical approach inspectable.
What to consider
- The result is a predicted decomposition, not recovery of the source artist's original layers or intent.
- Hidden content behind an object may be synthesized and should not be treated as factual evidence.
- The repository says text-to-multi-RGBA generation is limited because the released weights target decomposition.
- Prompts describe overall image content and cannot explicitly dictate the semantic content of each layer.
- Fine hair, transparency, reflections, shadows, typography and overlapping objects may produce artifacts or incorrect separation.
- The 20B BF16 model is resource-intensive compared with lightweight consumer editing tools.
- The authors report strong experimental performance, but production teams should benchmark their own image classes and manual-repair rate.
- An open model license does not grant rights to input images, trademarks, faces or generated edits; provenance and consent remain the user's responsibility.
Compare
Qwen Image Layered alternatives
The right alternative depends on the specific output, workflow, controls and budget your project requires.
Design
Adobe Firefly
A hosted creative suite for users who value integrated editing and commercial workflows over self-hosting an experimental layer model.
Explore Adobe Firefly →Design
Qwen-Image-3.0
A broader Qwen image-generation model when creating and rendering new images matters more than decomposing an existing raster into layers.
Explore Qwen-Image-3.0 →Questions
Qwen Image Layered FAQs
What does Qwen-Image-Layered do?
It takes a flat RGB image and predicts multiple transparent RGBA layers that can be independently moved, resized, recolored, removed or edited.
Is Qwen-Image-Layered free?
The official repository and weights are published under Apache 2.0 with no model-license fee listed. Users still pay for hardware, cloud compute, storage and engineering.
Can it recreate the original Photoshop file?
No. It predicts a useful layered representation; it cannot know the original layer names, masks, groups, adjustment settings, vectors, fonts or artist intent.
Can it export PSD files?
Yes. The repository's Gradio application supports PSD, PPTX and ZIP export, but the predicted layers may need manual cleanup.
How many layers can it create?
The model supports a variable requested count, and the official examples show different decompositions such as three and eight layers. Layers can also be recursively decomposed.
Can it generate new layered images from text?
The released weights accept text conditioning, but Qwen explicitly says text-to-multi-RGBA performance is limited because the model is fine-tuned for image decomposition.
What hardware does it need?
The official model card lists a 20B-parameter BF16 model and CUDA-oriented sample code. Exact memory and speed depend on resolution, layer count, steps, precision and implementation.
Are hidden pixels accurate?
Do not assume so. Content that was occluded in the source must be inferred, so it may look plausible while being invented.
Bottom line
Our Qwen Image Layered verdict
Qwen-Image-Layered is a compelling open building block for turning flat assets into editable approximations. It is most valuable to technical creative teams that can supply GPU infrastructure and manual QA; it should not be marketed as exact source-file recovery or factual reconstruction of hidden content.
Visit Qwen Image Layered website ↗Similar Design Tools
View all Design AI tools →
Stitch
Stitch - Google's experimental tool for turning ideas into UI designs for mobile and web applications
![FLUX.2 [klein]](https://cdn.prod.website-files.com/67ab272a36622522cbb3bac1/696d3ab90fa358b40187952c_Screenshot%202026-01-18%20at%201.55.31%E2%80%AFPM.png)
FLUX.2 [klein]
FLUX.2 Klein - BFL's new ultra-fast AI image editing model
.jpeg)
Nano Banana Pro
Nano Banana Pro - Google's image generation model with unprecedented control, improved text rendering and enhanced world knowledge.
Pencil (now pen.dev)
Pencil - An infinite design canvas for Claude Code
AI training for the future of work.
Get access to all our AI courses, hundreds of real-world AI use cases, live expert-led workshops, an exclusive network of AI early adopters, and more.
AI Courses
Get unlimited access to all of our current & upcoming industry-specific AI courses for the duration of your subscription.
Daily Guides
To keep up with the rapid pace of AI, our team publishes AI implementation guides daily. Our library contains 300+ practical use cases to automate real-world work.
Workshops
Join weekly, live, interactive sessions with industry leaders who are at the forefront of AI for hands-on implementation guidance and exclusive insights.
Community
Network with an exclusive community of AI-first professionals who are working smarter with AI. Learn how early adopters are using AI in their work and businesses.