LICA Dataset: Layered-Composition for Design

• LICA Layered-Composition Dataset is a comprehensive resource of over 1.5M annotated professional graphic designs spanning 20 sub-categories.
• It encodes structured design elements—including text, images, vectors, and groups—with extensive spatial, typographic, and temporal metadata.
• The dataset supports research in automated layout generation, constraint-preserving design editing, and temporally-aware synthesis using benchmark subsets.

The LICA (Layered Image Composition Annotations) Layered-Composition Dataset is a large-scale, richly annotated corpus of professional graphic designs structured as hierarchical, multi-element compositions. Encompassing over 1.5 million layouts across diverse design categories, LICA formalizes each design as a system of typed components—text, images, vectors, and groups—preserved with extensive per-element metadata. The dataset supports robust research in automated layout understanding, generative models for design, temporally-aware graphic composition, and structured editing, marking a substantial advancement over pixel-based datasets. Each composition is provided as rendered PNGs and in a detailed JSON schema that reflects both the graphical output and the underlying structural specification, with a particular emphasis on explicitly modeling design semantics and hierarchical relationships (Hirsch et al., 17 Mar 2026).

1. Dataset Scope and Composition Diversity

LICA comprises 1,550,244 discrete compositions ("layouts"), gathered to maximize structural and stylistic diversity as encountered in real-world graphic design workflows. Of these, 971,850 represent unique design templates, with 107,728 templates offering 2–24 style and content variants. This enables direct investigation into style consistency, template-conditioned generalization, and controlled design variation.

The dataset spans 20 professional sub-categories, with the majority of layouts reflecting high-volume social and business use-cases:

Category	Layouts	Percentage
Instagram Post	599,758	38.7%
Presentation	448,623	28.9%
Education	125,008	8.1%
Flyer	74,308	4.8%
Social Media	52,850	3.4%
Art Design	2,649	0.2%

The full range encompasses infographics, business cards, resumes, menus, posters, video advertisements, and more, establishing broad representativeness for layout and compositional research. The dataset does not specify a canonical train/validation/test split, encouraging user-defined strategies such as reserving entire templates as a held-out set to rigorously assess model generalization (Hirsch et al., 17 Mar 2026).

2. Annotation Schema and Data Model

LICA structures each layout as a hierarchical composition $C = \{e_1, e_2, \ldots, e_n\}$, with each element $e_i$ typed as text, image, vector (SVG or Lottie JSON), or group (including frame grids). Complex layouts are recursively decomposed into nested groups: $$C = G_1 \cup G_2 \cup \cdots \cup G_K, \quad G_j = \{e_{j1}, e_{j2}, \ldots, e_{j,m_j} \}$$ allowing for arbitrarily deep trees. While the released JSON flatten this hierarchy, parent indices and transform chains recover the original nesting.

Per-element metadata includes:

• Spatial geometry: $x_i, y_i$ (absolute top-left), $w_i, h_i$ (width, height)
• Transform chain: CSS-style encoding (rotation $\theta_i$, scale $s_{x_i}, s_{y_i}$, flips)
• Opacity: $\alpha_i \in [0,1]$
• Visibility: flag $v_i \in \{true, false\}$
• Layer order: $z_i \in \mathbb{N}$, governing stacking

Type-specific fields capture domain detail. For example, text components (8,146,222 total) preserve UTF-8 content, font family, size, weight, color, line height, letter spacing, text alignment, curvature (for arc text), auto-resize behavior, per-span style overrides, and (optionally) SVG background textures. Image (7,019,837) and vector (4,196,047, of which 619 are Lottie animations) elements retain original asset properties, clipping/cropping, overlays, and alternative text. Group components (5,315,918) encode child lists, group geometry, internal transforms, overflow, clip-paths, and frame-grid layout parameters for tiled arrangements.

3. Animated Layouts and Temporal Semantics

A distinctive aspect of LICA is its inclusion of 27,261 animated or "video" layouts. For each element $e_i$0 in a video composition, the temporal annotation $e_i$1 specifies:

$e_i$2

where $e_i$3 collects keyframes (timestamped property sets), $e_i$4 is an easing curve per segment, $e_i$5 the per-element duration, and $e_i$6 the start-time offset. Motion is categorized into 32 distinct types (e.g., slide-in, fade, scale, etc.), and temporal interpolation leverages CSS-standard easing (linear, ease-in, cubic-Bezier). For multi-slide animated designs, 11 inter-slide transitions are defined, each parameterized by direction and color. This temporal structuring enables research on generative video layouts and sequence modeling in design contexts (Hirsch et al., 17 Mar 2026).

4. Research Tasks and Benchmark Applications

LICA is structured to support and catalyze a spectrum of graphic design research problems:

• Structured layout generation: Predicting full component sets—including types, positions, visual styles, and layer order—conditioned on natural-language briefs or partial sketches.
• Constraint-preserving design editing (layer-aware inpainting): Training models to remove or substitute specific elements while maintaining coherent, plausible compositions.
• Controlled template variation: Utilizing template-variant groupings to learn consistency in style with content or palette swaps.
• Temporally-aware generation: Modeling and generating synchronized motion for compositional elements to produce animated layouts or story ads.
• Font classification and typographic synthesis: Generating per-character styling, font family/weight prediction, and curved text arrangement.
• Design quality evaluation and ranking: Learning metrics that combine geometry, typographic hierarchy, and template-conditional appropriateness.

At release, LICA proposes these new research tasks but does not yet provide quantitative baselines. A pilot subset comprising 1,000 layouts (including all data types and categories) is distributed for early benchmarking and comparative analysis (Hirsch et al., 17 Mar 2026).

5. Access, Contents, and Licensing

LICA is distributed through https://github.com/purvanshi-lica/lica-dataset. Release artifacts include PNG previews (for rendered appearance), hierarchical JSONs with full annotation for each layout, and the 1K-layout pilot benchmark subset. Licensing follows a non-commercial, Creative Commons-style agreement; detailed terms and correct citation form are documented in the repository. Proper citation is required for research use: Hirsch E., Yadav S., Garg M., Mehta P. “LICA: Layered Image Composition Annotations for Graphic Design Research,” 2024 (Hirsch et al., 17 Mar 2026).

6. Significance and Paradigm Shift in Graphic Design Datasets

LICA’s explicit modeling of layered composition, typographic detail, vector elements, and animation semantics marks a substantial departure from prior, pixel-centric graphics benchmarks. By encoding the structural and relational complexity of real-world designs, LICA encourages the development of models operating directly on compositional and semantic primitives rather than on rasterized appearance alone. This suggests a shift toward generative and interpretive systems with controllable, editable, and semantically grounded outputs. Its organization is tailored to foster novel directions in structured layout synthesis, layer-aware editing, temporal generation, and holistic design evaluation, potentially underpinning the next generation of vision-language and design-centric machine learning architectures (Hirsch et al., 17 Mar 2026).