DiagramBank: A Quality-Audited Dataset of Scientific Schematic Diagrams with Multi-Level Document Context

Introduction

Scientific papers use schematic diagrams to convey complex information. Existing datasets often mix different types of figures and lack contextual information. DiagramBank aims to provide a clear distinction by offering a dataset specifically for schematic diagrams, preserving both paper-level and figure-level context.

Dataset Construction and Quality Assurance

The dataset is constructed from PDFs sourced from OpenReview, with a focus on normalizing metadata and extracting figures and their contexts. A cascade-filtering approach is used to ensure high-quality diagram classification, distinguishing between diagrams and other visual materials.

Schematic classification

A core challenge is accurately classifying figures as diagrams. The classification process involves using a CLIP prefilter and a vision-language model to ensure precision in identifying schematic diagrams.

Dataset Statistics and Analysis

The paper provides quantitative breakdowns of the dataset, including the distribution of figure types and venue-level statistics, helping to contextualize the data for retrieval and authoring.

Scale and Subset Sizes

DiagramBank is derived from a large pool of figures, with a final release of 57,100 diagrams after a rigorous filtering process. The dataset includes various views based on confidence thresholds to cater to different user needs.

Figures Explained

The paper’s visual material highlights the workflow and the main system components.

• Figure 1 :: Figure 1: DiagramBank dataset and resource overview. This illustrative schematic contrasts weak figure-only extraction with context-rich records, a verified diagram collection, and dataset-grounded diagram-generation use. 2 Related Work and Resource Positioning Scientific-figure datasets provide important starting points, but they are not organized as diagramspecific resources with multi-level document context. SciCap (Hsu et al., 2021) supports caption generation over heterogeneous scientific figures; DocFigure (Jobin et al., 2019) focuses on figuretype classification; ACL-Fig (Karishma et al., 2023) covers ACL Anthology figures and tables; and visual-summary work ranks representative figures from paper content (Yamamoto et al., 2021; Zhong et al., 2025) . These resources show the value of figure-level supervision, but they either mix plots, tables, photos, and diagrams, omit the OpenReview-centered AI/ML venues studied here, or do not preserve the body-text context that explains how a schematic is used in the paper. Diagram-centric AI2D (Kembhavi et al., 2016) provides dense annotations, but its grade-school science domain differs from the conventions of modern AI/ML diagrams.Recent automated authoring and diagramgeneration systems, including AI Scientist (Lu et al., 2024) , autonomous paper-generation work (Yue et al., 2026) , Paper-Banana(Zhu et al., 2026a), AutoFigure(Zhu et al., 2026b), and diagram planners such as DiagrammerGPT and Sci-Doc2Diagrammer (Zala et al., 2023; Mondal et al., 2024) , increase the need for auditable scientific-.
• Figure 2: Figure 2(b) characterizes venue-level visual density using three indicators: figures per paper, visual breadth, and average caption length. Two patterns are particularly relevant for retrieval and diagram authoring.First, visual breadth is consistently high (79.8-97.6%), meaning that most papers contain at least one extractable figure. This is important for re-.
• Figure 2 :: Figure 2: Compact corpus statistics. (a) CLIP-predicted figure-type distribution over 452,339 extracted nontable figures, annotated with counts and mean confidence scores. (b) Venue-level visual density: bars show figures per paper, the line shows average caption length, and bar labels show breadth, defined as the percentage of papers with at least one extracted non-table figure.
• Figure 3 :: Figure 3: Example metadata-indexing use. Diagram-Bank metadata supports title retrieval, abstract reranking, and caption-level exemplar retrieval.
• Figure 4: Figure 4 reports descriptive diagnostics for the released corpus rather than task metrics. Panel (a) contextualizes caption-text retrieval, panel (b) explains why year-based analyses should account for venue coverage changes, and panel (c) shows why subject metadata is useful for domain-aware filtering.