Design39K: Layered Graphic Design Dataset

• Design39K is a layered graphic design dataset that decomposes raster visuals into editable foreground objects, background, and vectorized text.
• It supports diverse training modalities—such as text de-rendering, background-only rendition, and object removal—with 156,932 derived samples for robust supervision.
• The dataset underpins the Accordion framework, enabling top-down design planning with precise layer count, OCR, and attribute prediction performance.

Searching arXiv for the primary paper and closely related layered graphic design work. Design39K is an in-house, layered graphic design dataset introduced to train and evaluate Accordion, a top-down framework for converting raster graphic designs into editable layered compositions. It is designed around the professional graphic-design decomposition of a canvas into foreground objects, a background, and vectorized text, and provides supervision for structured design plans, raster-to-layer parsing, and text or object removal used in layered reconstruction. In the Accordion formulation, a design is modeled as $\{O_1, O_2, \ldots, O_N, B, T\}$, where $O_i$ denotes foreground object layers, $B$ the background layer, and $T$ vectorized text layer(s). The dataset was created to address the gap between visually strong but non-editable GenAI raster outputs and the layered, PSD/graphics-like representations required for practical editing workflows (Chen et al., 8 Jul 2025).

1. Definition, scope, and design rationale

Design39K is explicitly described as an in-house, layered graphic design dataset for top-down layered graphic design generation. Its primary purpose is to support three tightly coupled capabilities: learning structured design plans that encode layout and styles, training a vision-LLM (VLM) to parse raster designs into structured layer descriptions, and guiding text or object removal and de-rendering when converting flat designs into editable layered graphics (Chen et al., 8 Jul 2025).

The dataset is motivated by a specific mismatch in contemporary design pipelines. GenAI systems can readily produce visually rich graphic designs in pixel format, but those outputs are often non-editable. By contrast, professional workflows separate backgrounds, objects, and vectorized text into distinct layers. Design39K is intended to bridge that mismatch by supplying attribute-rich supervision over layered graphic elements rather than only flat raster exemplars. The dataset is therefore not merely a layout corpus; it is tailored to full graphic design decomposition, including background, foreground objects, and vectorized text with rich attributes.

Its content spans a diverse array of visual communication artifacts, including posters, book covers, advertisements, business cards, logos or brand materials, and various marketing and presentation designs. This breadth matters because the dataset is not restricted to a single design genre such as posters or document layouts. Instead, it reflects heterogeneous commercial and presentation-oriented graphic design practice.

A common misconception is to equate Design39K with a public benchmark. It is not. Design39K is the internal training and validation resource for Accordion, whereas benchmark evaluation is conducted on DesignIntention rather than on a held-out public test split within Design39K (Chen et al., 8 Jul 2025).

2. Dataset scale, splits, and layered representation

The dataset title corresponds directly to the approximately $39\text{k}$ base layered designs used for training. The reported splits and aggregate sample counts are as follows.

Split or count	Value
Training set	39,233 designs
Validation set	492 designs
Derived training instances	156,932 samples

Each base design is inherently layered. The dataset materializes separate layers for background, foreground objects, and vectorized text, in line with standard professional design practice. On average, each design contains approximately $1.02$ object layers and $3.11$ text regions (Chen et al., 8 Jul 2025).

The representation is built from layered design files, described as similar to PSD or AI files, from which the authors extract layer information including text, objects, backgrounds, and various attributes. For each design, the dataset includes raster views such as the full rendered design, a background-only image with text removed, and reference variants with nonsensical text created via inpainting. It also includes layer metadata used as VLM supervision: a structured list of elements, bounding boxes, and style attributes.

A central representation in Accordion is the design plan $\{D_*\}$, a sequence of dictionaries sorted bottom-to-top in stacking order. Each dictionary specifies attributes for a background segment, a foreground object, or a text element. Coordinates are normalized to $[0,336]$ because the reference image is resized so that its longer side is $336$ pixels before VLM ingestion. For text, paragraph-level annotations include bounding box, content, color, font, alignment, number of lines, and angle. RGBA values are originally represented in $O_i$0 and mapped to $O_i$1 for training, following COLE’s discretization strategy (Chen et al., 8 Jul 2025).

Object annotations provide bounding boxes, class information for planning, and implicit layer order through sequence position. Background annotations include a full-canvas bounding box and a description used in the planning stage when background is the only visible component. This makes the design plan a structured target for raster-to-layer prediction rather than a loosely specified captioning task.

3. Annotation schema and construction pipeline

Design39K is described as an in-house dataset collected and maintained by Adobe Research and collaborators, with no public license specified; its present status is internal and non-public (Chen et al., 8 Jul 2025). Each base design includes a title or description, separate layers for text, objects, and background, and extracted metadata derived from the original layered asset. Appendix E states that the title of each design is used as the description.

The construction pipeline expands each base design into multiple training modalities. First, the original layered design is rasterized and used for text de-rendering, where the input is a sensible raster design and the target is the original layered plan. Second, the authors generate a variant with nonsensical text by applying Stable Diffusion 1.5 inpainting, via ControlNet, to text regions. The inpainting strength is sampled uniformly from $O_i$2. The stated rationale is that strengths below $O_i$3 insufficiently alter the original text, whereas strengths above $O_i$4 overly destroy text and style and make training unstable. These inpainted images simulate AI-generated designs whose text is visually plausible but semantically unusable. Third, the authors remove all text layers to create a background-only reference for training text addition.

From each of the 39,233 base designs, the pipeline therefore yields one original-design sample, one nonsensical-text sample, and one background-only sample, supplemented by a questionnaire dataset used for quality selection in removal tasks. This produces 156,932 total training instances (Chen et al., 8 Jul 2025).

OCR is incorporated as auxiliary supervision. For original designs, PaddleOCR detection boxes and strings are included as supplemental input to the VLM during training. For GenAI-style references with nonsensical text, OCR boxes are retained as positional hints, but OCR strings are treated as unreliable; the target text content instead comes from the ground-truth description and layer annotations. This distinction is important because the dataset is designed to disentangle geometric text localization from textual correctness.

Preprocessing decisions are explicit. Images presented to the VLM are resized so the longest side is $O_i$5, with padding as needed for non-square layouts. Bounding boxes are normalized into $O_i$6. Output plans are serialized as text, with a mean output sequence length of $O_i$7 tokens and a maximum output length of $O_i$8 tokens, covering more than $O_i$9 of cases (Chen et al., 8 Jul 2025).

4. Role within the Accordion framework

Design39K is central to Accordion’s three-stage architecture, although its most direct role is in Stage 2, Design Planning. In that stage, the VLM receives a raster reference, resized to the $B$0-pixel scale, together with a combined prompt $B$1 as applicable, and predicts a serialized sequence of element dictionaries $B$2 derived from Design39K’s layer annotations (Chen et al., 8 Jul 2025).

The plan prediction target includes text bounding boxes, content, RGBA values, font, line count, alignment, and angle; object bounding boxes and layer order; and background bounding box plus a brief description when needed. Because the dictionaries are sorted bottom-to-top, the representation encodes not only element identity but also stacking order. This top-down decomposition is a defining difference from bottom-up methods such as COLE and Open-COLE, which gradually generate elements to create layered designs.

Stage 3 uses Design39K in a different way. The dataset supports text and object removal by supplying bounding boxes and layer counts, and it also provides questionnaire-style supervision for a VLM acting as a judge of removal quality. For each design and each removal task, the pipeline generates three different inpainting results in addition to ground truth, shuffles the four candidates, and trains the VLM to select the best result according to harmony or coherence, cleanliness of the background, and absence of artifacts or extra objects (Chen et al., 8 Jul 2025).

All 156,932 samples are used to train a single multi-task VLM, specifically LLaVA-1.5-7B with LoRA adaptation. The reported training configuration is 6 epochs with learning rate $B$3, using $B$4A100 $B$5 GPUs for approximately 36 hours. The same dataset therefore supports both plan prediction from raster references and result selection for removal quality (Chen et al., 8 Jul 2025).

5. Supported tasks and quantitative behavior

Design39K underpins all major tasks performed by Accordion. In text-to-template generation, the system takes a user intention and an AI-generated reference image containing scrambled or nonsensical text, then predicts the ground-truth design plan from the original layered design. In adding text to background, the input is a background-only raster plus caption, and the target is the original text layer configuration together with object layers. In text de-rendering, the input is the full rasterized design with OCR cues, and the target is the original text content, style, and positioning. These tasks together support the full conversion of flat designs into layered designs by predicting $B$6, removing text, segmenting and removing objects sequentially, and recombining extracted object layers, the final background, and rendered vectorized text (Chen et al., 8 Jul 2025).

Quantitative measurements on the Design39K evaluation set—the 492 validation samples—indicate that the dataset supports nontrivial attribute prediction. Reported layer-count errors are $B$7 L1 for text layers and $B$8 L1 for object layers. With OCR prompts, text detection F1 is approximately $B$9, OCR NED is approximately $T$0 at the paragraph level, and object detection F1 is approximately $T$1. Attribute-specific accuracies reported in ablations include color accuracy approximately $T$2, font accuracy approximately $T$3, line number accuracy approximately $T$4, alignment accuracy approximately $T$5, and angle accuracy approximately $T$6 (Chen et al., 8 Jul 2025).

These figures show that the dataset supports robust recovery of structural properties such as layer counts, localization, and text geometry, while finer appearance attributes such as color and font are more difficult. A plausible implication is that Design39K is especially strong as supervision for design planning and decomposition, while exact stylistic reconstruction remains a harder problem even with rich annotations.

The dataset also supports design variation workflows. Although those tasks involve additional upstream and downstream models, Design39K supplies the supervision needed for Accordion to parse and regenerate layers after image variation models distort text, and to preserve layout during operations such as translation (Chen et al., 8 Jul 2025).

6. Relation to adjacent resources, constraints, and future extensions

Design39K should be distinguished carefully from nearby resources. It is not DesignIntention: DesignIntention is the external benchmark from COLE, with 500 detailed text prompts across multiple design categories, while Design39K is the in-house layered dataset used for training (Chen et al., 8 Jul 2025). It also differs from the datasets behind COLE and Open-COLE in that it is explicitly layered and richly annotated at the level of text attributes, object layers, and background, and is tailored to top-down decomposition from raster to layers rather than primarily bottom-up generation.

Relative to other structured design corpora, Design39K occupies a particular position. The D3 search corpus of 7,860 deconstructed visualizations in “Searching the Visual Style and Structure of D3 Visualizations” is likewise a large structured design space, but its unit of analysis is the data visualization grammar of marks, encodings, axes, and style rather than editable layered graphic designs (Hoque et al., 2019). FashionEdit, introduced for iterative fashion design, is workflow-aware and edit-centric, but focuses on 2D garment images and multimodal refinement rather than layered decomposition of graphic designs (Li et al., 11 Sep 2025). Design39K is therefore distinctive in combining raster references, layered supervision, and typography-rich text attributes within a graphic-design setting.

Several constraints are dataset-related. The paper notes that object segmentation quality is limited in part by errors in VLM-predicted bounding boxes, with SAM achieving approximately $T$7 IoU on object extraction; this suggests that richer mask supervision such as polygon annotations could be beneficial in future versions (Chen et al., 8 Jul 2025). The text style vocabulary is limited to a predefined set of approximately 2,000 styles, which restricts reproduction of freeform lettering, special effects, and complex stylized typography outside the training set. The system also assumes that text layers are always on top of object layers, reflecting a simplifying bias in the dataset. Domain coverage is described as rich in English-centric commercial designs rather than explicitly multilingual. Finally, the non-public status of the dataset limits reproducibility and broader benchmarking.

The paper explicitly identifies future directions aligned with dataset expansion: extending the JSON schema in the planning stage to incorporate more attributes, such as richer typography and effects, and using the VLM as an agent that integrates more expert models. This suggests that Design39K is best understood not as a static benchmark artifact, but as a structured supervision substrate for editable graphic-design generation whose representational scope can be expanded as the underlying decomposition problem becomes more fine-grained (Chen et al., 8 Jul 2025).