One Image.
Every Layer.

Recover editable RGBA layers from a finished image.
Guide the decomposition with bounding boxes.

Try the Interactive Demo Explore the Layers
Original colorful raster design The Original
Independent RGBA Layers

Controllable Layer Decomposition for Reversible Multi-Layer Image Generation

37 layersOn a Single GPU
48 secAverage Decomposition
0.72Mask IoU · Predicted Boxes
−0.6%PSNR at 20% Box Noise
See CLD in action

From a finished image to editable layers

Watch the decomposition process, then explore every layer interactively below.

CLD decomposition demo Interactive walkthrough
One image in, a layer stack out

Peel a finished design apart

Hover to match boxes with layers. Drag the slider to separate the stack.

CLD - Decomposition Explorer
Input + bounding boxes

Decomposed stack
Layers
Abstract

CLD separates a flattened image into editable RGBA layers, guided by bounding boxes. LayerDecompose-DiT generates region-specific layers, while the Multi-Layer Conditional Adapter aligns them with the input image. Detector-predicted or user-defined boxes control the decomposition. A new benchmark evaluates quality, efficiency, and controllability. The recovered layers can be edited independently in tools such as PowerPoint.

Method

Bounding boxes guide each layer

Each box defines a layer. CLD generates within these regions to reduce computation.

CLD architecture: LD-DiT backbone with the MLCA control branch
(a) LD-DiT jointly denoises region-cropped layer tokens. (b) MLCA aligns image conditioning with each layer’s bounding box.
1

Region-aware tokens

Generate tokens only within each box, supporting up to 37 layers on one GPU.

2

Aligned conditioning

MLCA aligns image features with layer tokens to preserve spatial detail.

3

Dual-condition CFG

Keep image conditioning in both branches to reduce background colour leakage.

Two ways in

Automatic or user-guided

Compare detector-predicted boxes with user-defined boxes. Click a layer to enlarge it.

CLD - Box Source Comparison
Mode A · detector boxes automatic
Mode B · user boxes < 30 s to annotate

Robustness

Sloppy boxes are fine

Increase box noise from 0% to 20% to compare the recovered layers.

CLD - Bounding-Box Perturbation
Perturbed boxes
Resulting layers
Layer editing

Edit each layer independently

Edit the recovered layers in your browser, then export the composition as a PNG.

CLD Studio - Untitled.layers
Efficiency

Up to 37 layers on one GPU

Region-based generation reduces token usage. Compare inference time as layer count grows.

Inference Cost - Single GPU
Qwen-Image-Layered CLD (ours)
Avg layers11.0
Avg tokens10,986
Avg time47.6 s

On the 1,000-image PrismLayersPro test split, CLD reaches 37 layers. Qwen-Image-Layered runs out of memory at seven.

MethodBboxLayers RGB_L1 ↓IoU ↑Overall ↓ Semantic ↑Visual ↑Editability ↑
LayerD——0.05040.59450.22803.9783.8084.043
Sam2 + InpaintGT—0.06140.75460.15344.0153.6334.124
Qwen-Image-Layered—20.04350.63480.20444.0223.9204.054
Qwen-Image-Layered—30.02550.64010.19274.0193.9004.084
Qwen-Image-Layered—40.02650.59190.21734.0353.8784.101
Qwen-Image-Layered—50.03240.54340.24454.0433.8644.104
Qwen-Image-Layered—60.03980.49340.27324.0453.8594.110
CLD (ours)Pred—0.01640.72190.14724.0993.9134.162

PrismLayersPro test split. GT = ground-truth boxes; Pred = Qwen3-VL detector boxes. Reconstruction metrics: RGB_L1, IoU, Overall. Q-Insight scores: Semantic, Visual, Editability.