qwdwqfwq/topconf-paper-figure-gallery

🎨 Top-Conf Figure Gallery

顶会论文主图灵感画廊 · ICLR · ICML · NeurIPS · CVPR · ACL · AAAI · 2023–2026

See how well-designed papers lay out Figure 1 before you draw your own. A searchable, filterable gallery of Figure 1 / teaser figures. Static HTML/CSS/JS, no build, works offline.

画论文主图前，先看顶会论文怎么排版。 一个可搜索、可筛选的 Figure 1 / Teaser 画廊。纯静态、零构建、可离线打开。

🌐 Live gallery 在线画廊 · 🛠️ FigureForge 画图 · 🎬 Demo video 演示视频 · 📚 Methodology 数据与方法 · ⚖️ Image policy 版权 / 下架 · 🤝 Contributing 贡献指南

Top-Conf Figure Gallery collects Figure 1 / teaser figures (concept diagrams, system frameworks, pipelines, architectures, benchmark overviews) from six ML/AI conferences, 2023–2026. It is a design reference, not a paper ranking. Each card links to the source paper and records venue, year, authors and a visual-pattern tag.

• Filter by venue, year, acceptance tier and visual pattern; masonry layout with a lightbox.
• Oral, Spotlight and Best/Outstanding-paper figures carry gold / silver / red badges (ICLR, ICML, NeurIPS 2023–2026).
• Full-text search over titles and authors, e.g. DPO, robot, gaussian, agent.
• Every figure is hand-reviewed. Over 25 heuristic rules first remove default matplotlib charts, plain tables, GUI screenshots and unlabeled photo sets; a person then checks the rest page by page. See METHODOLOGY.md.
• Pure static site, no backend or build; clone and double-click.
• Open, reproducible pipeline: proceedings index → PDF download → Figure 1 crop → quality scoring → dHash de-duplication → manual review.

FigureForge drafts your Figure 1 from the gallery: describe your paper, and 3,452 top-conference figures provide the layout reference. No sign-up, no server; it runs in your browser. Try FigureForge →

▶ Full demo video, 1:58 with audio

1. Describe the paper. Paste the title and abstract, upload the PDF (the Step 1 model reads it and summarizes the sections matching overview / architecture / pipeline / results), or type one sentence giving the "input → method → output" story.

2. Pick a pattern. framework for system or multi-agent overviews, pipeline for end-to-end flows, architecture for internal model structure, conceptual for visual metaphors; also teaser and taxonomy.

3. Tick references. A CLIP + BM25 hybrid search ranks all 3,452 figures; the score is shown on each card. Tick at least 4 same-pattern cards for two-stage mode, or 2–3 for direct mode.

   

4. Two stages: summarize, then generate. A vision model reads the picks, distils their common layout, elements, palette and hierarchy, selects 2–3 representative figures, and only then generates.

   

5. Choose a mode, add your key, download. SVG mode gives accurate, editable text (Figma / draw.io / Illustrator); bitmap mode gives a publication-style layout to check in PPT / Figma. The key is stored only in your browser.

A bare prompt in a general chatbot returns a generic flowchart. FigureForge grounds the draft in top-conference figures, so it keeps the panel grid, comparative narrative and restrained palette of a paper figure.

• Runs entirely in your browser: the CLIP model and gallery index are bundled locally; paper content and keys stay on your device and the key goes only to the provider you choose.
• Bring your own key. Presets: Volcengine Ark, SiliconFlow, Zhipu, DeepSeek, OpenAI, Moonshot, Qwen, Tencent Hunyuan, Anthropic, and any OpenAI-compatible relay.
• The model list updates itself: the bundled manifest refreshes on load and the provider's live /models list is fetched once a key is entered. Anthropic has no list endpoint; Hunyuan and Anthropic block browser CORS, so use a relay for those.
• Scope: a high-quality first draft for the Figure 1 genre. Check text and numbers before submission.
• Open source forever, no paid version. MIT-licensed; the key pays the model provider directly. PRs are welcome for new papers, venues and label fixes, and the gallery is refreshed yearly with new proceedings.

Open a Suggest a figure issue with the paper link — no coding needed. For PRs, CONTRIBUTING.md gives the metadata format and quality bar. New venues such as CHI / RSS / CoRL are welcome.

Images belong to their authors and publishers; see IMAGES_POLICY.md for the educational-use and 72-hour takedown policy.

顶会论文主图灵感画廊收录六大 ML/AI 顶会 2023–2026 的 Figure 1 / Teaser：概念图、系统框架、流程图、架构图、任务全景。它是排版设计参考，不是论文排名。每张卡片链接原文，并记录会议、年份、作者和视觉模式标签。

• 按会议、年份、录取等级、视觉模式筛选；瀑布流布局，点击看灯箱大图。
• Oral / Spotlight / Best（杰出论文）主图带金、银、红角标（ICLR、ICML、NeurIPS 2023–2026）。
• 标题、作者、关键词全文搜索，如 DPO、robot、gaussian、agent。
• 每张图人工复核：25 条以上规则先剔除默认 matplotlib 图、纯表格、GUI 截图和无标签照片墙，再逐页人工检查。见 METHODOLOGY.md。
• 纯静态站点，无后端、无构建，克隆后双击即开。
• 管线全开源：会议索引 → PDF 下载 → Figure 1 裁剪 → 质量打分 → dHash 去重 → 人工复核。

FigureForge 基于画廊生成你的 Figure 1 初稿：描述论文，3,452 张顶会主图作为版式参考。免注册、无服务器，全部在浏览器本地运行。立即试用 →

▶ 完整演示视频，1 分 58 秒，有声音

1. 输入论文：粘贴标题和摘要、上传 PDF（步骤一的模型直接阅读，按概览、架构、流程、结果总结相关章节），或写一句话讲清「输入 → 方法 → 输出」。

2. 选模式：系统或多智能体总览选 framework，端到端流程选 pipeline，模型内部结构选 architecture，视觉隐喻选 conceptual；另有 teaser、taxonomy。

3. 勾选参考图：CLIP + BM25 混合检索全量 3,452 张图，卡片上显示相关度评分。两阶段模式勾至少 4 张同类型图（最多 10 张，越接近 10 张版式参考越充分），直接模式勾 2–4 张。

   

4. 两阶段：先归纳，再生成：视觉模型先读参考图，归纳共同的布局、元素、配色和层级，挑出 4–6 张代表图，再生成。

   

5. 选模式、填 Key、下载：SVG 模式文字准确、可用 Figma / draw.io / Illustrator 编辑；位图模式版式地道，在 PPT / Figma 核对即可。Key 只存在本机浏览器。

直接在通用对话框写 prompt，得到的是通用流程图。FigureForge 以顶会图为版式参考，初稿保留论文图的面板网格、对比叙事和克制配色。

• 全流程浏览器本地运行：CLIP 模型和画廊索引随仓库打包；论文内容和 Key 不离开设备，Key 只发往所选服务商。
• 自带 Key：预置火山方舟、硅基流动、智谱、DeepSeek、OpenAI、月之暗面、通义千问、腾讯混元、Anthropic，以及任意 OpenAI 兼容中转。
• 模型列表自动更新：打开时刷新内置清单，填入 Key 后拉取服务商实时 /models。Anthropic 无列表接口；混元、Anthropic 不支持浏览器跨域，这两家用中转。
• 能力边界：Figure 1 场景的高质量初稿，文字和数字提交前需人工核对。
• 纯开源、永不收费：MIT 许可，Key 费用直接付给模型厂商。欢迎提 PR 补论文、补会议、修标签，画廊每年随新论文集持续更新。

不会写代码也可以：开一个 Suggest a figure Issue，贴上论文链接。提 PR 前看 CONTRIBUTING.md，含元数据格式和收录标准。欢迎 CHI / RSS / CoRL 等新会议。

图片版权归原作者和出版方；教育用途与 72 小时下架政策见 IMAGES_POLICY.md。

ICLR / ICML 含 2023–2026；NeurIPS 含 2023–2025（2026 名单未公布）。

角标：红 = Best / Outstanding（含 Honorable Mention）、金 = Oral、银 = Spotlight。 口径以官方名单为准：ICLR 2023 的 notable top 5% 记为 Oral、top 25% 记为 Spotlight；ICML 2023 只有 OralPoster，记为 Oral；ICLR 2026 只设 Oral 与 Poster，无 Spotlight。

图片从正式出版 PDF 裁剪（约 180–216 DPI），网页版压缩为 ≤1500px JPEG，标注标题、全部作者和原文链接。各会议图量差异反映设计型主图的实际占比，与会议水平无关。

标签描述图的视觉功能，不是论文领域。

在线浏览 / Online: GitHub Pages。

本地运行 / Local server:

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

或直接双击 index.html（数据内联在 assets/figures.js，无需联网）。 Or double-click index.html; data is inlined in assets/figures.js.

├── index.html               # 画廊主页 gallery shell（搜索 / 筛选 / 灯箱）
├── assets/
│   ├── style.css / app.js   # 样式与交互
│   └── figures.js           # 图片元数据（由 data/figures.json 生成）
├── images/<venue>/final/    # 画廊图片 ≤1500px JPEG
├── data/
│   ├── figures.json         # 画廊唯一权威清单 single source of truth
│   └── pool/                # 全量候选论文池
├── scripts/                 # 数据管线，见 docs/METHODOLOGY.md
├── docs/                    # 方法文档、banner、演示
├── forge/                   # FigureForge：浏览器端检索与初稿生成
└── .github/workflows/       # GitHub Pages 自动部署

详见 docs/METHODOLOGY.md：

pip install -r requirements.txt
python scripts/build_pool.py          # ICLR / ICML / NeurIPS 候选池
python scripts/extract_all.py         # 下载 PDF、裁剪 Figure 1（断点续跑）
python scripts/build_pool_new.py cvpr,acl,aaai
python scripts/extract_new.py cvpr,acl,aaai 12
python scripts/score_select.py 980    # reject 规则、设计感评分、dHash 去重、选图
python scripts/assemble_gallery.py 1000   # 生成 figures.json / figures.js
python scripts/clean_stale.py         # 删除无引用图片
python scripts/enum_sheets.py iclr    # 40 格/页联系表，人工逐页复核
python scripts/sheet_qa.py 31         # 随机抽样复核

版本对照 / Versioning — releases 遵循 0.x semver：v0.3（六个会议，2,298 张，2026-09-20 发布）、v0.4（录取等级角标，2,730 张）、v0.5（2026 论文，3,528 张）、v0.6（FigureForge；去除 12 张跨代重复，3,516 张）。小红书按发帖计数：v1 = v0.3、v2 = v0.4、v3 = v0.6（v0.5 并入 v3 帖）。

• 60 张人工精选种子画廊（v0.1）
• ICLR / ICML / NeurIPS 自动管线扩量（v0.2–v0.3）
• CVPR / ACL / AAAI（2023–2025）同管线扩展（v0.3）
• Oral · Spotlight · Best Paper 等级索引与角标（v0.4）
• 2026 已公开会议主图与等级角标（v0.5）
• FigureForge：浏览器端 CLIP+BM25 检索、自有 API 生成初稿（v0.6）
• CVPR / ACL / AAAI 的 Oral / Highlight 等级索引
• NeurIPS 2026 录用公布后补全（预计 2026 年 12 月）
• 持续人工复核，开放社区 PR 补图
• CHI / CoRL / RSS / EMNLP 等会议扩展
• 相似图推荐、按配色检索、一键导出 BibTeX

欢迎补图、修标签、修 bug、扩展新会议。不会写代码可直接开 Suggest a figure Issue，贴上论文链接。提 PR 前读 CONTRIBUTING.md，注意收录标准：不收纯照片墙和默认图表（折线图 / 柱状图 / 散点图 / 热力图这类「脚本就能画出来」的图会在收录时被 scripts/filter_charts.py 筛掉）。

感谢以下社区贡献者（按贡献采纳时间排序），他们的实现均已合入当前版本：

Thanks to @timelic and @TansyZenix — their lightbox transitions, SVG sandboxing, keyboard accessibility and the data-validation CI all ship in the current release. Contributions are credited in CHANGELOG.md as well; PRs for new papers, venues and label fixes are always welcome.

图片版权归原作者与出版方，本仓库仅作非商业教育、研究用途的署名索引。ICLR / ICML(PMLR) / NeurIPS / CVPR(CVF) / ACL 正式出版论文多为 CC BY 4.0；AAAI 论文为 fair-use 学术参考。如需下架，请提 Issue 或邮件 939123836@qq.com，核实后 72 小时内删除。详见 IMAGES_POLICY.md。

@misc{topconf_figure_gallery,
  title  = {Top-Conf Figure Gallery: A Curated Gallery of Figure 1 / Teaser Designs from ICLR, ICML, NeurIPS, CVPR, ACL and AAAI (2023-2026)},
  year   = {2026},
  url    = {https://github.com/qwdwqfwq/topconf-paper-figure-gallery}
}

每会议-年份评分最高的 2 张；完整 3,452 张见在线画廊。

ICLR

2023 · NTFields: Neural Time Fields for Physics-Informed Robot Motion Planning	2023 · Learning Heterogeneous Interaction Strengths by Trajectory Prediction with Graph Neural Net…	2024 · Online Stabilization of Spiking Neural Networks
2024 · Unleashing Large-Scale Video Generative Pre-training for Visual Robot Manipulation	2025 · GraphArena: Evaluating and Exploring Large Language Models on Graph Computation	2025 · Open-World Reinforcement Learning over Long Short-Term Imagination

ICML

2023 · Neural Latent Aligner: Cross-trial Alignment for Learning Representations of Complex, Natur…	2023 · AudioLDM: Text-to-Audio Generation with Latent Diffusion Models	2024 · Bridging Environments and Language with Rendering Functions and Vision-Language Models
2024 · Recurrent Early Exits for Federated Learning with Heterogeneous Clients	2025 · Locate-then-edit for Multi-hop Factual Recall under Knowledge Editing	2025 · R*: Efficient Reward Design via Reward Structure Evolution and Parameter Alignment Optimiza…

NeurIPS

2023 · SwiftSage: A Generative Agent with Fast and Slow Thinking for Complex Interactive Tasks	2023 · LayoutPrompter: Awaken the Design Ability of Large Language Models	2024 · A survey and benchmark of high-dimensional Bayesian optimization of discrete sequences
2024 · Task Me Anything	2025 · SingRef6D: Monocular Novel Object Pose Estimation with a Single RGB Reference	2025 · RepoMaster: Autonomous Exploration and Understanding of GitHub Repositories for Complex Tas…

CVPR

2023 · TrojDiff: Trojan Attacks on Diffusion Models With Diverse Targets	2023 · Detecting and Grounding Multi-Modal Media Manipulation	2024 · Learning Large-Factor EM Image Super-Resolution with Generative Priors
2024 · VBench: Comprehensive Benchmark Suite for Video Generative Models	2025 · <a href="https://openaccess.thecvf.com/content/CVPR2025/html/Jagpal_EIDT-V_Exploiting_Intersections_in_Diffusion_Trajectories_for_Model-Agnostic_Zero-Shot_T…	2025 · <a href="https://openaccess.thecvf.com/content/CVPR2025/html/Wang_FSFM_A_Generalizable_Face_Security_Foundation_Model_via_Self-Supervised_Facial_Representat…

ACL

2023 · C on FEDE : Contrastive Feature Decomposition for Multimodal Sentiment Analysis	2023 · Prosody- TTS : Improving Prosody with Masked Autoencoder and Conditional Diffusion Model Fo…	2024 · DELL : Generating Reactions and Explanations for LLM -Based Misinformation Detection
2024 · G rounding GPT : Language Enhanced Multi-modal Grounding Model	2025 · N eg VQA : Can Vision Language Models Understand Negation?	2025 · A Troublemaker with Contagious Jailbreak Makes Chaos in Honest Towns

AAAI

2023 · MDM: Molecular Diffusion Model for 3D Molecule Generation	2023 · What Does Your Face Sound Like? 3D Face Shape towards Voice	2024 · PosDiffNet: Positional Neural Diffusion for Point Cloud Registration in a Large Field of Vi…
2024 · Wavelet Dynamic Selection Network for Inertial Sensor Signal Enhancement	2025 · EditBoard: Towards a Comprehensive Evaluation Benchmark for Text-Based Video Editing Models	2025 · Enhancing Multivariate Time-Series Domain Adaptation via Contrastive Frequency Graph Discov…

_{图片版权归原作者及出版方所有，点击图片跳转原文。}