LICA: Layered Image Composition Annotations for Graphic Design Research

This dataset is a curated sample of 1,148 multi-layer graphic design compositions from the LICA dataset, spanning 19 design categories.

LICA provides hierarchical component-level representations of graphic designs, going beyond pixel-based approaches to capture the full compositional structure: text, images, vectors, and groups with spatial, typographic, and stylistic metadata.

Dataset Details

Dataset Description

Each layout is represented as a hierarchy of typed components (TEXT, IMAGE, GROUP, vector elements) with CSS-like positioning and styling properties. The dataset includes:

• 1,148 layouts sampled across categories and media types
• Per-layout metadata (category, dimensions, template grouping)
• Component-level layout specifications (position, size, typography, colors, transforms)
• Semantic annotations (descriptions, aesthetics, tags, inferred user intent) at both layout and template level
• Animated layouts with per-component keyframes and motion parameters (for video-category entries)

The full LICA dataset contains 1,550,244 layouts across 971,850 unique templates. This repository provides a representative sample for evaluation purposes.

• Curated by: Lica World
• Language(s): English
• License: Creative Commons Attribution 4.0 International (CC BY 4.0)

Dataset Sources

• Paper: LICA: Layered Image Composition Annotations for Graphic Design Research
• Repository: github.com/lica-world/lica-dataset

Uses

Direct Use

This dataset is designed for evaluation and benchmarking of graphic design models. Supported research tasks include:

• Structured layout generation -- predicting complete design compositions from briefs and metadata
• Constraint-preserving design editing -- localized editing while maintaining layout coherence
• Font classification and typographic analysis -- leveraging per-character style annotations
• Template-consistent layout variation -- controlled generation from multi-variant template groupings
• Short-form marketing video generation -- using per-component keyframe metadata
• Design quality evaluation and ranking -- developing learned metrics for design assessment

Dataset Structure

lica-data/
├── metadata.csv                              # per-layout metadata
├── layouts/
│   └── <template_id>/
│       └── <layout_id>.json                  # component-level layout spec
├── images/
│   └── <template_id>/
│       └── <layout_id>.png or .mp4           # rendered layout (image or video)
└── annotations/
    ├── template_annotations.json             # template-level annotations
    └── <template_id>/
        └── <layout_id>.json                  # per-layout annotation

metadata.csv

Category Distribution (this sample)

Layout JSON (layouts/<template_id>/<layout_id>.json)

Each layout file is a flat JSON object with the canvas specification and a list of components:

{
  "components": [ ... ],
  "background": "rgb(252, 252, 252)",
  "width": "1920px",
  "height": "1080px",
  "duration": 3
}

Component types

Each component has a type field and CSS-like positioning/visual properties directly on the object.

TEXT -- positioned text element

{
  "type": "TEXT",
  "text": "Hello World",
  "left": "108px", "top": "200px", "width": "400px", "height": "50px",
  "color": "rgb(255, 255, 255)",
  "fontSize": "48px",
  "fontFamily": "League Spartan--400",
  "fontWeight": "400",
  "textAlign": "center",
  "lineHeight": "52px",
  "letterSpacing": "0em",
  "textTransform": "none",
  "fontStyle": "normal",
  "transform": "none"
}

IMAGE -- positioned image

{
  "type": "IMAGE",
  "src": "https://storage.googleapis.com/lica-video/<uuid>.png",
  "left": "0px", "top": "0px", "width": "1920px", "height": "1080px",
  "transform": "none",
  "opacity": 1,
  "overflow": "hidden"
}

GROUP -- container/shape element with optional clip path

{
  "type": "GROUP",
  "left": "108px", "top": "463px", "width": "555px", "height": "508px",
  "background": "rgb(255, 255, 255)",
  "backgroundColor": "rgb(255, 255, 255)",
  "clipPath": "path(\"M0,0 ...\")",
  "transform": "none"
}

Annotations

Per-layout (annotations/<template_id>/<layout_id>.json):

{
  "description": "Visual description of the specific layout",
  "aesthetics": "Notes on design style, composition, visual hierarchy",
  "tags": "comma, separated, keyword, tags",
  "user_intent": "Inferred purpose or goal of the design",
  "raw": "Concatenation of all fields above"
}

Template-level (annotations/template_annotations.json):

A JSON object keyed by template UUID. Each entry has the same fields (description, aesthetics, tags, user_intent, raw) but describes the overall design theme shared by all layouts in the template.

Dataset Creation

See the LICA paper for full details on the data collection methodology.

Layouts were created by third-party annotators using a custom rendering engine that represents each design as a hierarchy of editable components. Design intent prompts were generated using large language models to ensure diverse coverage across categories. Quality assurance combined automated structural validation with manual review for visual fidelity and adherence to design guidelines.

Personal and Sensitive Information

The dataset consists of professionally designed graphic layout templates. No personal, sensitive, or private information is contained in the data.

Bias, Risks, and Limitations

• The category distribution is skewed toward Instagram Posts and Presentations, which together make up the majority of the sample.
• Good design is inherently context-dependent: a bold asymmetric layout effective for a music festival poster may be inappropriate for a corporate report. Universal quality metrics derived from this data should be interpreted with caution.
• This sample of ~1,148 layouts is not representative of the full dataset's scale (1.55M layouts).

Citation

@article{Hirsch2026LICA,
  title   = {LICA: Layered Image Composition Annotations for Graphic Design Research},
  author  = {Hirsch, Elad and Yadav, Shubham and Garg, Mohit and Mehta, Purvanshi},
  journal = {arXiv preprint arXiv:2603.16098},
  year    = {2026}
}

Dataset Card Contact

Purvanshi Mehta (hello@lica.world)