3,516 Top-Conference Figure 1 Examples: A Searchable Gallery for Paper Illustration Inspiration

Writing a paper often stalls at Figure 1: the main illustration must convey the entire contribution at a glance, yet there is no standard template. The Top-Conf Figure Gallery solves this by providing a curated, searchable collection of Figure 1 images from ICLR, ICML, NeurIPS, CVPR, ACL, and AAAI (2023–2026).

What's Inside

The gallery contains 3,516 figures distributed as follows:

ICLR: 745

ICML: 748

NeurIPS: 966

CVPR: 308

ACL: 384

AAAI: 301

The lower CVPR count is explained: many CVPR Figure 1s are qualitative result walls or video frame strips, which were filtered out because they offer little layout reference. NeurIPS leads because its papers frequently use system-overview diagrams.

Each figure card shows conference, year, authors, original paper link, and a visual-mode tag. Filters include conference, year, acceptance level (Oral/Spotlight/Best with gold/silver/red badges), and visual mode. Full-text search works on title, author, and keywords (e.g., DPO, robot, gaussian, agent).

Visual Modes: What the Figure Is Doing

The most useful classification is the visual-mode tag, which describes the figure's visual function, not the paper's domain: conceptual — visual metaphor for core concept (e.g., DPO, Tree of Thoughts) framework — system or multi-agent module overview (e.g., HuggingGPT, MetaGPT) pipeline — end-to-end staged data flow architecture — model internal layers and tensor connections taxonomy — task, capability, or data classification panorama teaser — designed mixed text-image hero visual

If you need a multi-agent system overview, filter framework to see how peers arrange panels.

Curation Pipeline: Not Just Crawled

Figures were not simply scraped. A multi-stage pipeline ensured quality:

Machine pre-filter with 25+ rules to exclude: pure image collages, default matplotlib charts, unlabeled coordinate plots, text walls, full-page tables, software screenshots, result photo walls.

Perceptual hashing for deduplication.

Human page-by-page review: all candidate figures were arranged in 40-per-page contact sheets (81 pages total). Every page was inspected, removing ~1,400 unqualified crops.

The authors explain why human review was necessary: random sampling showed unqualified crops appeared across all score bands, so truncating low scores alone would not guarantee quality.

FigureForge: From Reference to Draft

FigureForge uses the gallery as a reference library for AI-assisted figure generation. The workflow:

Describe your paper — paste title and abstract, upload PDF (model reads overview, architecture, process, results), or write a one-sentence "input → method → output" summary.

Choose mode — framework for system overview, pipeline for end-to-end flow, architecture for model structure, conceptual for visual metaphor.

Select references — CLIP + BM25 hybrid retrieval over 3,500+ figures; each card shows relevance score. Two-stage mode requires ≥4 same-type references; direct mode 2–4.

Induce then generate — vision model first reads selected references, induces common layout, elements, color, hierarchy, picks representative exemplars, then generates the draft.

Download format — SVG (editable text, importable into Figma/draw.io/Illustrator) or bitmap (faithful layout for PPT/Figma verification).

The entire process runs locally in the browser; CLIP model and gallery index are bundled with the repo. Paper content and API keys never leave the device. Keys are sent only to the chosen provider (pre-configured: Volcano Ark, SiliconFlow, Zhipu, DeepSeek, OpenAI, Moonshot, Qwen, Tencent Hunyuan, Anthropic, plus any OpenAI-compatible proxy). The authors emphasize output is a high-quality draft for Figure 1 scenarios; text and numbers must be manually verified before submission.

How to Use

Easiest: open the GitHub Pages deployment. For local offline use:

git clone https://github.com/qwdwqfwq/topconf-paper-figure-gallery.git
cd topconf-paper-figure-gallery
python -m http.server 8000

Or simply double-click index.html — all data is inlined in assets/figures.js, no network required.

The data pipeline is also open source: conference index → PDF download → Figure 1 crop → quality scoring → deduplication → human review. Scripts are in scripts/ for extending to other conferences.

Limitations and Copyright

Images cannot be reused directly; copyright belongs to original authors and publishers. The repo provides attribution index (title, all authors, paper link) for educational/research use. A 72-hour takedown policy honors rights-holder requests.

Figure count differences do not reflect conference quality; CVPR's low count stems from its result-wall Figure 1s being filtered out.

FigureForge is beta: output is a draft, not final. Text/numbers need manual check. Hunyuan and Anthropic require a proxy due to browser CORS restrictions.

Core Value

Academia lacks a path to surface good design. Top-conference Figure 1s often represent days of polishing, yet that design experience stays buried in proceedings. This project makes 3,500+ figures searchable, tagged, and categorized — turning tacit design knowledge into retrievable material. Additionally, the open-sourced pipeline demonstrates how to systematically collect and curate design assets, a reference for research-tool builders.

GitHub:

https://github.com/qwdwqfwq/topconf-paper-figure-gallery