3,452 Paper Figures You Can Actually Search, Filter, and Open Offline
You're staring at a blank canvas, trying to sketch out Figure 1 for your paper. You know what you want to say, but how do you lay it out? What do the best papers in your field actually do with their teaser figures? You could scroll through PDFs one by one, but that's a miserable way to spend an afternoon. The topconf-paper-figure-gallery takes a different approach: it's a searchable, filterable collection of 3,452 Figure 1 and teaser images pulled from top conference papers, built as a pure static site you can open offline.
What It Does
At its core, this is a gallery of paper main figures from ICLR, ICML, NeurIPS, CVPR, ACL, and AAAI, spanning 2023 through 2026. The collection includes 26 Best, Oral, and Spotlight papers alongside 616 and 950 from other tiers (the badges break this down as 26, 616, and 950). You can search and filter through the whole set, then open any figure in a large preview.
The project is deliberately simple on the technical side. It's a pure static site with zero build step—no bundler, no framework, no npm install ritual. That means you can clone it and open it locally, or just visit the hosted version. The repo also lists Python 3.10+ as a requirement, which points to the data collection and processing side of things rather than the viewer itself.
There's also a companion piece called FigureForge, which lives at a /forge path on the same site. The README shows four examples of figures it generated—a robot policy paper, a controllable video generation paper, a reed straw slurry study, and an LLM sycophancy analysis—and notes that exports are editable down to all the details. It's a separate tool aimed at actually producing figures rather than just browsing them.
Why It's Cool
-
It solves the "where do I even start" problem. Most people writing a paper have looked at plenty of figures, but not systematically. Having thousands of teaser images in one searchable place turns a vague aesthetic question into something you can actually study.
-
Static and offline-first is the right call. A gallery like this doesn't need a backend. Because it's zero-build and pure static, it's trivially forkable, hostable, and archivable. You can download it and have the whole thing working on a plane. That's a genuinely thoughtful design decision for a reference tool.
-
The conference and year filters matter more than they sound. Figure conventions differ across venues—what flies at CVPR isn't what you'd draw for ACL. Being able to narrow to your target conference and recent years is the difference between inspiration and noise.
-
The licensing is handled sensibly. The code is MIT, and the images are CC BY with attribution, with a separate
IMAGES_POLICY.mdspelling out the details. That's the kind of thing projects often gloss over, and it's good to see it addressed up front. -
FigureForge is a natural extension. Browsing examples is useful; being able to generate your own editable figure is the logical next step. The examples shown span robotics, video generation, materials science, and LLM analysis, which suggests the tool isn't narrowly scoped to one subfield.
One honest caveat: the README is clearly written for a Chinese-speaking audience first, with translations into English, French, Russian, Korean, and Japanese. The English version exists, but if you're relying on the main README you'll be reading through a translation layer.
How to Try It
The fastest path is the hosted gallery—no setup at all:
- Open the online gallery and start searching or filtering by venue and year.
- If you want to generate figures rather than just browse them, head to the FigureForge tool.
To run it locally (and keep it available offline), clone the repo:
git clone https://github.com/qwdwqfwq/topconf-paper-figure-gallery.git
cd topconf-paper-figure-gallery
Since it's a zero-build static site, you can open the entry HTML directly in a browser. If you're touching the Python side of things, install the dependencies first:
pip install -r requirements.txt
Contributions are welcome—there's a CONTRIBUTING.md and the repo explicitly invites PRs, paper recommendations, and code. If you're in China, there's also a mirror on AtomGit.
Final Thoughts
This isn't a tool that will transform your workflow, and it doesn't pretend to be. It's a well-organized reference that saves you from the tedious work of hunting through PDFs when you need layout inspiration. It's best suited to grad students and researchers who are actively writing papers and want to see how strong work in their venue structures its main figure—plus anyone curious enough to try FigureForge for generating their own. The static, offline-friendly design and clear licensing make it easy to trust and easy to keep around. If you're about to sketch Figure 1, it's worth a look before you open your drawing tool.
Follow @githubprojects for more developer tools and open source projects.