🎨

Image Edit - QWEN-Layered

🛠

Turn a Flat Image into Editable Photoshop-Style Layers

Decomposes an image into editable RGBA layers (foreground objects, background, text, etc.) — like turning a flat photo into a Photoshop file where each element sits on its own transparent layer. Pair with List-Extract → QWEN-Image-Edit → List-Inject → Layer-Compose to edit a single layer (change color, replace an object, rewrite text) and put it back without touching anything else in the image.

What it does

Qwen-Image-Layered takes a normal flat image and automatically separates it into multiple transparent layers, exactly like the layers in Photoshop or Figma. Each layer holds one distinct element (a person, an object, the background, a text block, etc.) with a transparent RGBA background so it doesn't cover the other layers. Once the image is split like this, you can edit any single layer without disturbing the rest of the image — recolor an object, resize it, move it, delete it, or replace it with something else — then recompose the layers back into a final image. The key win is consistency: because the edit only touches one layer, everything else in the image stays pixel-identical. That's often impossible with normal image editing models, which tend to subtly change unrelated parts of the image whenever you edit one thing.

Problem it solves

  • Edit one thing, keep everything else exactly the same – The traditional pain point with AI image editing is that "just change the shirt color" often ends up changing the face, background, or details too. Layered decomposition physically isolates the edit to a single layer
  • Precise object operations – Move, resize, recolor, or delete a single object cleanly with no repainting or ghosting
  • Text editing without touching art – Isolate the text block onto its own layer, edit just the text, and recompose — surrounding artwork stays intact
  • Object replacement – Swap "girl → boy" or "cat → dog" by editing just that layer (typically with QWEN-Image-Edit) and re-inserting
  • Clean object removal – Delete an unwanted object cleanly, since it's on its own layer with transparency behind it
  • Flexible layer count – Not locked to a fixed number of layers — decompose into 3, 4, 8, or more as the image demands
  • Recursive / infinite decomposition – Any layer can itself be further decomposed if you need finer control
  • Photoshop-like editability from any photo – Get an editable layered file from an image that was never authored in layers

Input/Output

  • Input:
    • Image (required) — a single source image to decompose
      • Format: PNG / JPG / JPEG / WebP (RGBA or converted to RGBA on load)
      • Any aspect ratio — the output aspect follows the input
    • Text (optional) — a short caption or hint describing the image's content, to guide how the model splits it into layers
      • Examples: "portrait of a woman in a blue blazer, indoor office background", "product shot of a red sneaker on a white surface with brand text"
      • When left empty, the model auto-captions the image internally (via its built-in vision-language understanding). Providing your own caption can improve decomposition on unusual images or when you want the model to prioritize a specific element
      • English and Chinese are both supported
        • image
  • Options:
    • Number of layers to decompose into (default 4)
      • How many separate layers the image should be split into (variable — pick what suits the image)
    • Working resolution / quality (default 640)
      • Output aspect follows the input; 640 is the recommended bucket for this version (1024 also supported)
    • Diffusion steps (10 – 60, default 50)
      • More steps = higher quality, slower
    • Random seed (-1 = random, default 1)
      • Set a fixed integer to reproduce the exact same decomposition
    • Advanced Options:
      • CFG scale (1 – 10, default 4.0) — how strictly the model follows the decomposition intent
      • Negative prompt (default empty) — what to avoid in the output; leave empty for none
  • Output: A list of N RGBA layer images
    • Each layer is a transparent PNG containing one element of the original image
    • Stacked in order, they recompose to (approximately) the original image
    • Ready to hand off to List-Extract for downstream editing
      • image
      image

How to Use

The studio node is designed to sit at the front of a layered edit workflow:

  1. QWEN-Layered — decompose the input image into N RGBA layers
  2. List-Extract — pull out the single layer you want to edit
  3. QWEN-Image-Edit — apply your edit to that one layer (recolor, replace, rewrite text, etc.)
  4. List-Inject — put the edited layer back into the layer list
  5. Layer-Compose — recompose the full layered stack into a single final image

Accuracy & Speed

Performance

Model type
Diffusion-based image-to-layers decomposer
Base model
Qwen/Qwen-Image (fine-tuned for layered decomposition)
Recommended resolution
640 (this version); 1024 also supported as an alternate bucket
Default layer count
4 (freely adjustable — 3, 8, or more supported)
Recursive decomposition
Any output layer can itself be re-decomposed for finer isolation
Languages
English and Chinese prompts
Precision
bfloat16 recommended

Technical Details

Task
Image-Text-to-Image (specifically: layer decomposition into RGBA stack)
Base model
Qwen/Qwen-Image
Diffusers pipeline
QwenImageLayeredPipeline
Default Inference Hyperparameters
num_inference_steps=50 (range 10 – 60), true_cfg_scale=4.0 (range 1 – 10), layers=4, resolution=640, cfg_normalize=True, use_en_prompt=True
Text Encoder
Qwen2.5-VL (via transformers>=4.51.3)
Supported Frameworks
Hugging Face Diffusers (main branch — pip install git+https://github.com/huggingface/diffusers), Safetensors
Publication
"Qwen-Image-Layered: Towards Inherent Editability via Layer Decomposition" — arXiv:2512.15603 (Yin et al., Dec 17, 2025)

Compliance & Provenance

Provider
Alibaba (open weights)
Provider type
Specialized
License
Apache 2.0
EU AI Act risk class
Limited Risk
Art. 50 transparency
Required — outputs are marked. See AI Policy §2.
Region availability
Available globally
Training data summary
Provider model card / training data

For more on how we classify models and mark outputs, see our AI Policy.

Model Source