Does Synthetic Layered Design Data Benefit Layered Design Decomposition?

Technical Summary: Synthetic Layered Design Data for Layered Design Decomposition

Problem Statement
Recent advances in image generation have enabled the production of high-quality images, yet these outputs are inherently flattened raster images where foreground elements, backgrounds, and text are entangled. This "last-mile" gap hinders flexible post-generation editing, as modifying specific components often requires manual effort or unstable generative editing that fails to preserve geometry and semantics. While graphic design decomposition aims to recover editable RGBA layers from raster images, existing approaches face scalability challenges. Current methods rely on scarce proprietary layered assets (e.g., Crello) or partially synthetic datasets (e.g., PrismLayersPro) with limited sample sizes (typically ~20K). These datasets are orders of magnitude smaller than those used in modern generative modeling, contradicting established scaling laws and limiting model robustness and generalization.

Methodology
The authors propose SynLayers, a fully synthetic dataset designed to train effective layer decomposition models without relying on real-world layered assets. The methodology is built upon CLD (Controllable Layer Decomposition), a state-of-the-art framework that uses a crop-then-denoise pipeline with bounding-box guidance.

1. Synthetic Data Construction (SynLayers):

   • Multi-Source Composition: The pipeline assembles diverse elements from multiple sources: base designs from PrismLayersPro, RGB/RGBA foreground layers, and text layers.
   • Augmentation: Additional elements are introduced probabilistically, including LAION image crops, rendered text layers, and AlphaVAE foreground objects.
   • Spatial Arrangement: A low-overlap algorithm places layers by sampling candidate positions and selecting the one that minimizes normalized overlap with existing elements, ensuring spatial coherence and layout diversity.
   • Caption Construction: Initial captions are generated by traversing a 3x3 spatial grid and describing layer content based on source metadata. These raw descriptions are refined by a Vision-Language Model (VLM) to produce coherent, fluent whole-image captions.

2. Automated Inference Pipeline:

   • To eliminate reliance on manual annotations during inference, the authors train a VLM (fine-tuned Qwen3-VL-8B-Instruct) to jointly predict the whole-image caption and layer bounding boxes from a single raster image.
   • This VLM acts as an upstream detector, providing the structured inputs (captions and boxes) required by the downstream CLD decomposition model.

Key Contributions

• SynLayers Dataset: A scalable, fully synthetic dataset comprising up to 500K samples with ground-truth RGBA layers, bounding boxes, and structured captions.
• Data-Centric Investigation: A comprehensive study demonstrating that purely synthetic data can serve as a viable, scalable alternative to non-synthetic or partially synthetic datasets for graphic design decomposition.
• Automated Pipeline: An end-to-end system where a VLM automatically generates the necessary conditioning inputs (captions and boxes) for the decomposition model, removing the need for manual annotation at inference time.
• Balanced Distribution Control: The synthetic pipeline allows for precise control over layer-count distributions, addressing the severe class imbalance (e.g., scarcity of high-layer-count samples) found in real-world datasets.

Experimental Results
The study evaluates the SynLayers-trained model against baselines trained on PrismLayersPro and the Qwen-Image-Layered model.

• Performance vs. Real Data: With an equivalent training budget (18K samples), the SynLayers-trained model outperforms the PrismLayersPro baseline across key metrics, including Layer PSNR (+1.01) and Composite PSNR (+0.83).
• Scaling Laws: Performance improves with data scale but exhibits diminishing returns. Optimal results for Layer FID (5.97) and Composite FID (10.35) are observed at medium scales (20K–30K), with performance stabilizing beyond 50K samples.
• Robustness to Layer Count: SynLayers-trained models demonstrate superior robustness across varying layer counts (1 to 35+). In challenging high-layer scenarios (13–35 layers), Mask IoU improved from 0.901 to 0.910, and Composite PSNR increased from 29.48 to 30.25 compared to the baseline.
• Generalization: On an out-of-distribution real-world test set (147 images), the SynLayers model achieved higher composite fidelity (PSNR 29.35, FID 35.40) compared to the PrismLayersPro baseline (PSNR 28.74, FID 44.23) and significantly outperformed Qwen-Image-Layered.
• Qualitative Improvements: Visual comparisons show sharper typography, cleaner object boundaries, and fewer visual artifacts (e.g., color bleeding) in the SynLayers-trained outputs.

Significance and Claims
The paper claims that synthetic layered design data provides a practical and scalable foundation for graphic design decomposition. By demonstrating that purely synthetic data can match or exceed the performance of models trained on scarce real-world assets, the work addresses the bottleneck of data availability in this domain. The authors posit that this approach enables the construction of more reliable and controllable graphic design editing systems, bridging the gap between image generation and user-editable outputs. The study encourages the broader adoption of synthetic data as a substitute for proprietary assets, particularly for tasks requiring fine-grained control over layered structures.