Does Synthetic Layered Design Data Benefit Layered Design Decomposition?
:::info Stub — Full Engineering Breakdown Coming This paper has a linked code implementation and was featured on Hugging Face Papers with 6 upvotes. A full breakdown with production viability rating, implementation notes, and honest limitations is being written. Subscribe to AI Letters → :::
| Authors | Kam Man Wu et al. |
| Year | 2026 |
| HF Upvotes | 6 |
| arXiv | 2605.15167 |
| Download | |
| Code | https://github.com/YangHaolin0526/SynLayers |
Abstract
Recent advances in image generation have made it easy to produce high-quality images. However, these outputs are inherently flattened, entangling foreground elements, background, and text within a fixed canvas. As a result, flexible post-generation editing remains challenging, revealing a clear last-mile gap toward practical usability. Existing approaches either rely on scarce proprietary layered assets or construct partially synthetic data from limited structural priors. However, both strategies face fundamental challenges in scalability. In this work, we investigate whether pure synthetic layered data can improve graphic design decomposition. We make the assumption that, in graphic design, effective decomposition does not require modeling inter-layer dependencies as precisely as in natural-image composition, since design elements are often intentionally arranged as modular and semantically separable components. Concretely, we conduct a data-centric study based on CLD baseline, which is a state-of-the-art layer decomposition framework. Based on the baseline, we construct our own synthetic dataset, SynLayers, generate textual supervision using vision language models, and automate inference inputs with VLM-predicted bounding boxes. Our study reveals three key findings: (1) even training with purely synthetic data can outperform non-scalable alternatives such as the widely used PrismLayersPro dataset, demonstrating its viability as a scalable and effective substitute; (2) performance consistently improves with increased training data scale, while gains begin to saturate at around 50K samples; and (3) synthetic data enables balanced control over layer-count distributions, avoiding the layer-count imbalance commonly observed in real-world datasets. We hope this data-centric study encourages broader adoption of synthetic data as a practical foundation for layered design editing systems.
Engineering Breakdown
Plain English
This paper investigates whether purely synthetic layered design data can train models to decompose flat generated images back into editable layers (foreground, background, text, etc.). The core finding is that effective graphic design decomposition doesn't require modeling complex inter-layer dependencies the way natural image decomposition does, meaning synthetic training data alone can be a scalable alternative to scarce proprietary layered assets.
Key Engineering Insight
Graphic design decomposition is fundamentally simpler than natural scene understanding—you don't need to learn how layers physically interact or occlude each other. This means synthetic data generation becomes viable at scale, since you're not trying to capture the physics-like dependencies that plague synthetic data in other domains.
Why It Matters for Engineers
Image generation tools produce flat outputs that are locked down and hard to edit. For any production system letting users modify generated designs (ads, social content, layouts), you need to decompose those flat images into layers. This paper shows you can do that with synthetic training data, eliminating the bottleneck of collecting expensive labeled design assets and making it feasible to ship design-decomposition features.
Research Context
Prior work either relied on rare proprietary layer datasets or tried to bootstrap from limited structural priors. This created a scalability ceiling. By proving that purely synthetic data works for graphic design—unlike in natural image tasks where domain gap is critical—the paper opens a path to practical decomposition systems that can scale with synthetic data generation, not proprietary asset libraries.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::