cs.CVMar 17, 2026
SaveLICA: Layered Image Composition Annotations for Graphic Design Research
Abstract
We introduce LICA (Layered Image Composition Annotations), a large scale dataset of 1,550,244 multi-layer graphic design compositions designed to advance structured understanding and generation of graphic layouts. In addition to rendered PNG images, LICA represents each design as a hierarchical composition of typed components including text, image, vector, and group elements, each paired with rich per-element metadata such as spatial geometry, typographic attributes, opacity, and visibility. The dataset spans 20 design categories and 971,850 unique templates, providing broad coverage of real-world design structures. We further introduce graphic design video as a new and largely unexplored challenge for current vision-language models through 27,261 animated layouts annotated with per-component keyframes and motion parameters. Beyond scale, LICA establishes a new paradigm of research tasks for graphic design, enabling structured investigations into problems such as layer-aware inpainting, structured layout generation, controlled design editing, and temporally-aware generative modeling. By representing design as a system of compositional layers and relationships, the dataset supports research on models that operate directly on design structure rather than pixels alone.
Figures & tables
Figure 1: LICA samples. Rendered design layouts spanning diverse categories and aspect ratios, with corresponding component-level bounding-box annotations extracted from the structured layout JSON. Unlike prior datasets that provide only coarse bounding boxes, LICA encodes the full component hierarchy with per-element typographic, spatial, and style metadata.
| Component Type | Total Count | Avg. per Layout |
|---|---|---|
| Text | 8,146,222 | 5.25 |
| Image | 7,019,837 | 4.53 |
| Group | 5,315,918 | 3.43 |
| Vector | 4,196,047 | 2.71 |
| Total | 24,678,024 | 15.91 |
Table 1: Component type distribution across the 1,548,444 layouts in LICA, totaling over 24.6 million individually annotated elements with an average of 15.91 components per layout.
| # | Category | Layouts | % |
|---|---|---|---|
| 1 | Instagram Post | 599,758 | 38.7 |
| 2 | Presentation | 448,623 | 28.9 |
| 3 | Education | 125,008 | 8.1 |
| 4 | Flyer | 74,308 | 4.8 |
| 5 | Social Media | 52,850 | 3.4 |
| 6 | Business Card | 40,220 | 2.6 |
Table 2: The 20 design categories in LICA, spanning 1,550,244 layouts in total.
Figure 2: Layout semantic metadata. Each layout is annotated with multiple levels of semantic information beyond structural attributes, covering layout description , aesthetic analysis , user intent and component descriptions . Together, these annotations bridge the gap between low-level structural representation and high-level design semantics.
Figure 3: Template variants in LICA. Each row shows four layout variants derived from the same design template. Variants share a coherent design language – color palette, typographic system, visual motifs, and compositional logic, while differing in element positioning, content, imagery, and spatial arrangement. This structure captures a nuanced form of design knowledge: the variants are unified not by identical parameters but by an implicit style identity that governs how elements relate to one another, a property that is often difficult to articulate verbally yet immediately recognisable to a trained designer. LICA contains 107,728 templates with multiple variants (up to 24 per template), providing large-scale natural supervision for studying style adherence, controlled variation, and content-agnostic layout consistency.
Figure 4: Template and layout semantic metadata. Each layout in the dataset is associated with language annotations describing the layout, user intent, and aesthetic analysis (examples of user intents are shown at the bottom). For templates, an additional annotation is provided at the template level (top).
| Video Category | Count |
|---|---|
| Social Media Videos | 16,275 |
| Presentations | 6,576 |
| Other (e.g., collages, messages) | 4,406 |
| Total | 27,257 |
Table 3: Distribution of animated and video design layouts in LICA. Unlike static layouts, these compositions encode temporal structure directly within the component hierarchy.
Figure 5: Graphic design video in LICA. Each row shows four keyframes sampled from an animated design composition. Unlike static layouts, these compositions encode per-component temporal structure, including easing curves, start-time offsets, and durations, directly within the component hierarchy. Animations operate at the semantic layer level: each text block, image, or shape carries its own entrance, transition, and exit behavior independently of other elements, and timing is used as a deliberate compositional tool. This temporal logic is fundamentally different from natural video, where motion arises from continuous physical dynamics. In design video, motion is sparse, discrete, and driven by communicative intent rather than physical plausibility, making it a distinct and under-explored domain for temporal generative modeling. LICA contains 27,261 such animated layouts spanning presentations, social media videos, mobile ads, and other temporal design surfaces.
Figure 6: Data collection platform. (a) The generation interface allows annotators to import design assets, inspect individual layers, and edit component parameters (position, typography, color, animation timing) on an interactive canvas with real-time rendering. (b) The verification pipeline enables the QA team to review annotated layouts through a systematic approval workflow combining automated structural checks with manual visual inspection.
| Dataset | Size | Representation | Component Types | Style Metadata | Animation |
|---|---|---|---|---|---|
| Magazine [ 4 ] | 4K | Bounding boxes | 6 coarse | No | No |
| Crello [ 3 ] | 24K | Element attributes | 4 coarse | Partial | No |
| CGL-Dataset [ 2 ] | 60K | Bounding boxes | 4 coarse | No | No |
| PKU-PosterLayout [ 1 ] | 10K | Bounding boxes | 3 coarse | No | No |
| LICA (ours) | 1.5M | Full hierarchy | 4 coarse + subtypes | Yes | Yes |
Table 4: Comparison of LICA with existing graphic design datasets. LICA provides an order-of-magnitude increase in scale, full compositional hierarchy with typed components, rich per-element style metadata, and, uniquely, temporal animation annotations.
| Property | Unit / Format |
|---|---|
| Font family | String (e.g., Arimo , Alfa Slab One ) |
| Font size | Pixels |
| Font weight | Numeric (100–900) |
| Text color | Hex (e.g., #0A0203 ) |
| Line height | Pixels or multiplier (e.g., 120% ) |
| Text alignment | left | center | right | justify |
Table 5: Element-level typographic properties annotated for each text component in the dataset.
Figure 7: Rich text annotations. (a) Style ranges allow individual characters or spans within a single text element to carry distinct properties, e.g., colors and style (underline), capturing the mixed style typography common in professional design. (b) The curvature parameter governs arc-shaped text paths, enabling non-linear text layouts that are frequent in logos, badges, and decorative headings. The presented images are cropped from layouts.
Figure 8: Image transform and clipping annotations. (a) Arbitrary rotations are encoded as part of the component transform chain, which differs across the heart images (e.g., two rotated instances of the same heart, marked by orange arrows). (b) Non-rectangular clip paths are stored explicitly: here, the original rectangular image is cropped to a circle (orange arrow indicates the clipped region), using a two-level representation that separates the visible frame from the full uncropped asset. (c) Optional stroke overlays reference separate SVG assets composited on top of the image as decorative borders: here, the white rhombus shape (orange arrow) is a stroke overlay layered above the underlying photograph. (b) and (c) are cropped from full layouts. ((b) and (c) are cropped from layouts)
Figure 9: Lottie animation components. Two Lottie-format animated elements, each shown at two timeframes. These self-contained JSON animation assets carry their own keyframes and easing definitions independently of the layout’s temporal parameters. The presented images are cropped from layouts.
Figure 10: Graph and stroke SVG components. (a) Graph SVG elements encode data visualisations and infographic overlays. (b) Stroke SVG elements (marked by orange arrows) capture decorative vector strokes and brushstroke overlays. Together, these subtypes account for over 1.3 million components in the dataset. The images are cropped from layouts.
Figure 11: Category-aware layout generation. Given a textual design brief a generative model predicts a complete design composition, jointly placing elements such as shapes, images, and text with their spatial configuration, typographic attributes, and layer ordering. The final output is an overlay of multiple layers.