LICA: A Dataset of 1.5M Layered Graphic Design Compositions

TL;DR

Lica World (formerly Anything.design) is releasing LICA — Layered Image Composition Annotations — a public dataset of 1,550,244 multi-layer graphic-design compositions across 20 categories, each represented as a typed hierarchy of text / image / vector / group components with full per-element spatial geometry, typographic, opacity, and visibility metadata. 27,261 of the compositions are animated layouts with per-component keyframes, introducing graphic design video as a new data modality. The blog positions LICA as an order-of-magnitude jump over prior graphic-design datasets and the substrate for Lica World’s own models.

Key claims

1,550,244 multi-layer compositions across 20 design categories, drawn from 971,850 unique templates [Stats table].
Each composition is a typed component hierarchy with four component types — text, image, vector, group — and rich per-element metadata: spatial geometry, typographic properties (font, weight, size), opacity, visibility [§Intro].
27,261 animated layouts with per-component keyframes and motion parameters, intended to enable temporally-aware layout generation, motion-consistent editing, and animation-parameter prediction [§A New Modality].
LICA is positioned as “an order of magnitude leap” over existing graphic-design datasets in scale, and as the training-data foundation for the models Lica World is building [§What’s Next].
The intended downstream tasks are explicitly enumerated: layer-aware inpainting, structured layout generation, temporally-aware generative modeling, constraint-aware editing, and temporally grounded design modeling [TL;DR, §What’s Next].

Method

LICA is a dataset, not a model — the post does not describe its construction pipeline beyond noting that compositions are paired with the original user intent that generated each template (visible in the per-category sample browser). Compositions are stored as structured “LICA configs” alongside rendered composites, suggesting a serialized scene-graph format (rather than rasterized layers) that preserves editability. Animated subset (~1.8% of the dataset) stores per-component keyframes and motion parameters as a structured timeline. No license, access procedure, evaluation split, or VLM/LLM-in-the-loop curation detail is disclosed in this announcement.

Results

This is a release announcement and reports no model results. Reported headline numbers are dataset statistics: 1.55M compositions / 971,850 templates / 27,261 animations / 20 categories / 4 component types.

Why it’s interesting

LICA is a candidate public successor to ART’s private 1M-sample MLTD dataset, which is currently the unstated quality bar for layered-image generation (ART: Anonymous Region Transformer for Variable Multi-Layer Transparent Image Generation §3.4 reports 1M graphic designs with K=2–50 layers). At ~1.55M compositions with full typed hierarchy and per-element metadata, LICA is plausibly large enough and structured enough to train an ART-style anonymous-region transformer on. It is also strictly larger and richer than LASAGNA-48K (Controllable Layered Image Generation for Real-World Editing (LASAGNA)), which is built by running LayerDecomp (Generative Image Layer Decomposition with Visual Effects (LayerDecomp)) over photographic imagery and VLM-filtering — a fundamentally different (decomposed-from-photos) data source. The animated-layout subset is the more novel angle: no existing layered-image dataset surfaced on the wiki ships per-component keyframes, and this is the substrate a video-domain analogue of ART or OmnimatteZero: Fast Training-free Omnimatte with Pre-trained Video Diffusion Models would need. Whether the typed-scene-graph format ports well to the gray-background-encoded RGBA latent representation that ART/LayerDecomp converged on is open; vector and text components in particular are not trivially RGBA layers.