LogoBrief-10K — three-minute narration
01 — A dataset about design
Most logo datasets say where a mark appears or which brand it represents. LogoBrief-10K describes how each mark is designed: ten thousand exact original SVGs, paired raster renders, and structured design annotations.
02 — What is inside
Each row includes an SVG path, raster image, identity fields, and one of eighty-seven categories. Every row also has open-vocabulary style tags, a one-sentence motif, and a generated brief covering typography, color, composition, and shape.
03 — Geometry and embeddings
Where available, text boxes locate lettering in source-pixel coordinates for typographic evaluation. Every logo also has a unit-normalized, float-sixteen DINOv3 ViT-L sixteen embedding: one thousand twenty-four pooled visual features aligned by row.
04 — Not just famous brands
A naive intersection placed sixty-one percent of examples in the top popularity decile. LogoBrief samples roughly eighty-eight thousand candidates across twenty Tranco bins and an unranked stratum, reducing the final share to eleven point three percent.
05 — A reproducible split
The release has nine thousand and six training rows, plus four hundred ninety-seven each for validation and test. Splits follow the popularity strata. There is one render per domain, and no identity crosses a split.
06 — Fresh opt-out review
Before publication, every domain was checked across robots dot text, TDMRep, AI text, LLMs text, response headers, and homepage metadata. Unclear cases were excluded. A second review covered prose that deterministic matching could not interpret.
07 — Audit evidence, not a claim
The audit removed twenty-six point three percent of the first draw, then backfilled by stratum. The release documents ten thousand accepted domains and two thousand seven hundred seventy-two exclusions, with reasons. Rights holders can request removal without justification.
08 — Known limitations
Annotations are model-generated, not human-verified. Object descriptions can be wrong, sparse fields vary with popularity, and DINOv3 has a natural-image prior. Screening replaced eighty-two blank renders. An unreliable palette field was removed.
09 — Responsible access
The images are third-party trademarks; LogoLabs claims no rights in them and grants none. Access is gated, manually reviewed, and limited to non-commercial research. Users cannot redistribute the data and must apply published removals.
10 — What it enables
LogoBrief-10K supports image-to-text research, retrieval, feature analysis, text-geometry evaluation, and representation studies. It is compact, richly annotated, and explicit about construction and limitations: a dataset for asking better design questions without ignoring provenance.