Source-linked AI summary
Does Synthetic Layered Design Data Benefit Layered Design Decomposition?
Kam Man Wu, Haolin Yang, Qingyu Chen, Yihu Tang, Jingye Chen, Qifeng Chen
TL;DR
Flattened generated images make flexible post-generation editing difficult, while existing layered-design supervision is scarce or only partially synthetic. The paper builds SynLayers around CLD with synthetic layer composition, VLM-generated captions, and predicted boxes, finding improved reconstruction quality, scalable training behavior, and robust layer-count coverage. Its evaluation remains bounded by unmodeled professional effects, rectangular-box detection, limited external validation, and absent repeated-seed uncertainty estimates.
Problem
Flattened raster outputs entangle design elements, while existing layered datasets rely on scarce proprietary assets or partially synthetic data with limited scalability.
Method
The paper constructs SynLayers through multi-source synthetic layer composition, adds VLM-generated semantic supervision, and automates decomposition inputs with VLM-predicted boxes using CLD.
Results
Synthetic data improves reconstruction quality over the comparison baseline and supports robust performance across layer-count settings, with gains increasing through intermediate training scales before stabilizing.
Takeaways & Limitations
SynLayers provides a practical and scalable foundation for graphic-design decomposition and controllable design systems.
Takeaways & Limitations
The evaluation does not model complex professional design effects, relies on rectangular-box detection, lacks repeated-seed uncertainty estimates, and uses limited out-of-distribution validation without layer-level ground truth.
Abstract
from arXiv · showhide
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.
1 Introduction
Graphic design decomposition addresses the difficulty of editing flattened raster outputs by recovering editable layers. This study proposes SynLayers and finds that synthetic data can match or exceed alternatives, though scaling is non-monotonic and synthetic data supports balanced layer-count evaluation.
- Motivation: Flattened raster outputs entangle foregrounds, backgrounds, and text, making subsequent edits manual and motivating recovery of editable RGBA layers.Generative editing may fail to preserve unchanged-region geometry and semantics.
- Motivation: Existing decomposition datasets rely on limited real data or partially synthetic data with layouts derived from real designs, creating scalability constraints.Crello and PrismLayersPro are each described as having around 20K samples.
- Approach: SynLayers combines synthetic layer composition with VLM-generated captions to provide semantic supervision for layer decomposition.The study builds on the CLD layer decomposition method and uses diverse synthetic samples.
- Findings: At 18K training samples, synthetic data improves layer PSNR by +1.01 and composite PSNR by +0.83 over the comparison baseline.The passage reports comparable or better performance on most metrics, indicating stronger reconstruction at layer and image levels.
- Findings: Performance does not improve monotonically with more synthetic data; Layer FID reaches 5.97 at 20K and Composite FID reaches 10.35 at 30K.The best results in these examples occur at medium training scales.
- Findings: Synthetic data improves robustness across layer-count settings, including Mask IoU from 0.901 to 0.910 and Composite PSNR from 29.48 to 30.25 for 13–35 layers.The study presents synthetic data as a practical and scalable option for training layer-decomposition models.
2 Related Works
Prior work highlights scarce and insufficient layered graphic-design supervision, while synthetic data offers a route to address data scarcity across vision tasks. SynLayers applies this strategy to fully synthetic layered design decomposition.
- Layer decomposition: LayerD uses iterative matting and inpainting but provides limited user control, illustrating limitations among existing decomposition approaches.The passage situates LayerD among research efforts focused on layered image decomposition.
- Layered design datasets: Layered graphic-design datasets are limited because clean RGBA layers usually remain in proprietary design files rather than public raster datasets.Existing resources may provide annotations or simple composites but lack the complexity and rich text needed for graphic-design tasks.
- Synthetic data: Synthetic data is used broadly to mitigate scarce or expensive real-world supervision in modern computer-vision pipelines.The related work cites applications spanning multiple vision domains.
- SynLayers: SynLayers constructs fully synthetic layered data by recombining multi-source assets into composite designs, boxes, and captions.The construction overview includes complete supervision for decomposition training.
- Synthetic data: Synthetic images have achieved parity with real data in recognition, while synthetic assets have supported 3D reconstruction-model training.These examples motivate investigating synthetic data for graphic-design decomposition.
3 Methodology
SynLayers builds fully synthetic layered designs through multi-source composition, caption construction, and automated VLM-predicted inference inputs for CLD. Its placement procedure reduces overlap while preserving layer aspect ratios and promoting layout diversity.
- Preliminary Knowledge of CLD: CLD crops features within each layer's bounding box, jointly denoises the cropped regions, and decodes them into RGBA layers.This restricts computation to valid layer extents and supports varying layer counts.
- Multi-Source Synthetic Dataset Construction: SynLayers constructs fully synthetic training samples by combining layers from multiple sources into diverse graphic designs.The pipeline includes reference layouts, cross-source foreground integration, and auxiliary elements such as images, text, and foreground objects.
- Layer Composition and Placement: Candidate sampling places each layer where normalized overlap with existing elements is minimized.Zero-overlap candidates are preferred; otherwise, the candidate with the least overlap is selected.
- Layer Composition and Placement: The placement procedure preserves layer aspect ratios while promoting spatial separation and layout diversity.
- Caption Construction and Refinement: Captions are constructed by assigning layers to a 3 × 3 spatial grid and describing regions in reading order.Each layer combines its spatial location with a source-specific description.
- VLM-Guided Automated Input Generation: A single VLM predicts the global caption and foreground boxes from the raster image to supply CLD's inference inputs automatically.The predicted structured outputs directly condition the downstream decomposition model.
4 Experimental Results
SynLayers improves decomposition quality over PrismLayersPro and Qwen-Image-Layered in matched-scale, qualitative, OOD, and layer-count evaluations. Medium-scale training provides the strongest trade-off, while longer training yields diminishing or unstable gains.
- Qualitative Results: SynLayers produces sharper boundaries, cleaner text separation, and fewer artifacts than the PrismLayersPro baseline in qualitative comparisons.The model also more closely matches ground-truth layer layouts in real-world samples.
- Quantitative Results: Layer PSNR improves from 26.22 to 27.23 and Composite SSIM from 0.944 to 0.950 when comparing matched 18K PrismLayersPro and SynLayers training.The matched-size comparison supports synthetic layers as a data-efficient substitute.
- Scaling Analysis: Layer FID reaches 5.97 at 20K and Composite FID reaches 10.35 at 30K, with performance stabilizing and diminishing returns beyond medium scales.The scaling analysis spans dataset sizes from 1K to 500K.
- OOD Generalization: The SynLayers checkpoint improves OOD composite PSNR from 28.74 to 29.35 and FID from 44.23 to 35.40 against original CLD on 147 real-world images.Inference inputs are automatically generated by the Qwen3-VL detector, and layer-level ground truth is unavailable.
- Training Dynamics: From 10K to 40K, Composite PSNR rises from 29.24 to 30.89, while FID peaks at 12.45 before degrading and fluctuating through 500K.The reported 40K–60K window balances pixel fidelity and distributional alignment.
- Layer-Count Robustness: SynLayers improves most decomposition-oriented metrics across all four layer-count bins, showing benefits beyond a narrow complexity range.The evaluation uses bins spanning 1 to 35 ground-truth layers; SynLayers remains concentrated in moderate-complexity samples rather than uniformly covering counts.
5 Conclusion
The paper proposes SynLayers as a synthetic layered-design dataset and combines it with VLM-based supervision and automated inference inputs. The results support synthetic data as a practical, scalable foundation for graphic design decomposition and controllable editing systems.
- SynLayers is proposed as a synthetic layered-design dataset for graphic design decomposition.
- A VLM-based detector is integrated with the CLD backbone to automate inference inputs and support layer reconstruction evaluation.
- Synthetic supervision consistently enhances layer reconstruction quality and composite-image visual effects within the evaluated system.
A.1 Additional Methodology Details
The methodology preserves the CLD backbone while building scalable synthetic layered data, refining captions with VLMs, and serializing layer metadata for decomposition and detector training.
- Inherited CLD backbone: The inherited CLD architecture is unchanged, while supervision and conditioning are modified for layered inputs.LD-DiT crops each layer to its quantized bounding box, and the backbone remains largely frozen during adaptation.
- Dataset construction: SynLayers is generated on 1024 × 1024 RGBA canvases by combining base designs, donor layers, and auxiliary content.The pipeline samples and removes foreground layers, adds donor designs, and probabilistically inserts LAION crops, rendered text, or AlphaVAE objects.
- Serialization and inference: Each sample stores layer-wise source metadata, bounding boxes, captions, image paths, and a composite image for downstream training.The detector is trained on serialized whole-caption and box targets, while inference converts raster images into these structured inputs.
- Caption supervision: Raw grid-based captions are refined by Qwen2.5-VL-3B-Instruct into coherent layer-aware text supervision.The refinement prompt requests a holistic overview followed by layer details, overlaps, and visible English text.
A.4 Training Details for detector
The detector training setup fine-tunes Qwen3-VL-8B-Instruct with LoRA on spatially annotated SynLayers samples at high image resolution.
- Detector training: Qwen3-VL-8B-Instruct is fine-tuned with LoRA on SynLayers to predict whole-image captions and bounding boxes.Inputs use 1024 × 1024 images to support high-resolution spatial awareness.
- Detector training: The detector training configuration is summarized in Table 6.The supplied table passage identifies the hyperparameter reference but does not include the parameter-value rows.
A.5 Training Details for the Decomposition Model
The decomposition model is fine-tuned from the released CLD setup while preserving frozen backbone weights and adding trainable LoRA modules.
- Model initialization: The model is initialized from FLUX.1[dev], the released CLD adapter checkpoint, and released CLD LoRA weights.New LoRA modules are injected into LD-DiT and MLCA, while layer-position embeddings are also trainable.
- Evaluation setup: Figure 9 compares SynLayers-trained inference outputs with ground-truth layered designs.The comparison presents cases from the decomposition-model evaluation.
A.7 Quantitative Evaluation of the Qwen3-VL detector
The Qwen3-VL detector is evaluated on object discovery, box localization, and caption quality using strict detection metrics and GPT-4.1-based semantic judging.
- Bounding-box evaluation: 91.26% precision, 82.34% recall, and 86.57% F1 at IoU=0.50 are reported under strict object-discovery evaluation.This protocol penalizes both missed and hallucinated boxes; the reported counts are 1525 true positives, 146 false positives, and 327 false negatives.
- Bounding-box evaluation: Matched-box metrics evaluate localization quality only for successfully matched bounding boxes, whereas strict metrics include discovery errors.The evaluation distinction is stated in the Table 8 description and detector analysis.
- Evaluation interpretation: The evaluation framework is designed to show that object discovery is strong but still imperfect in a fully automated pipeline.The stated motivation is to provide a more realistic assessment than matched-box metrics alone.
- Automated input generation: The detector is prompted to return a JSON object containing one whole-image caption and bounding boxes for visible layers or objects.Each box uses the format [x0, y0, x1, y1] for a 1024 × 1024 image.
A.8 Compute Resources
The main decomposition-model training runs used four NVIDIA H800 GPUs, while detector training used four NVIDIA H20 GPUs. One main decomposition sweep checkpoint required roughly five days of wall-clock time, alongside additional exploratory and failed runs.
- 4 NVIDIA H800 GPUs supported the main decomposition-model training runs.
- Roughly 5 days of wall-clock time were required to obtain one checkpoint for the main decomposition sweep.
- Detector training used 4 NVIDIA H20 GPUs, with additional exploratory runs and failed tests beyond the reported checkpoints.
- The study targets more accessible editable-layer research for creative, educational, research, and design-assistance workflows.
C Limitations
The study’s synthetic pipeline and evaluation remain limited in modeling complex professional design effects, handling irregular or overlapping elements, and validating downstream real-world editing performance.
- The synthetic pipeline does not yet model complex blending modes, stylized transparency, or highly irregular text layouts.
- The VLM-based detector relies on rectangular boxes and JSON generation, limiting performance on non-rectangular or highly overlapping design elements.
- Detector evaluation covers box metrics and whole-caption quality, but not full downstream human-editing studies.
- Evaluation lacks repeated-seed uncertainty estimates and uses a small out-of-distribution real-world set without layer-level ground truth.