🎨 Top-Conf Figure Gallery

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

✨ FigureForge 生图成品 · Generated examples

以下是画廊 FigureForge 的实际生图示例。导出可编辑所有细节。

机器人策略与执行 _{Robot policy & execution}	层级工作记忆 _{Hierarchical working memory}
工作记忆系统 · 从输入到结果 _{Memory system overview}	LLM 迎合行为因果分析 _{Internal causes of sycophancy in LLMs}

Last commit

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

🎨 算力赞助伙伴｜Infistar.cc 无限星河

感谢 Infistar.cc 无限星河 为本项目提供大模型 API 算力支持！

🔬 主流模型一站接入：聚合 Claude、OpenAI、DeepSeek 等模型，支持图表代码生成（TikZ / Python）、论文润色与图表解析。
📉 降低科研成本：部分模型价格低至官方 0.1 折，倍率以平台实时公示为准。
⚡ 高并发支持：支持开票；模型可用性与限流规则以平台实时状态为准。
🎁 专属体验礼：通过专属推广链接注册，可领取 5 元体验额度与首充特惠，具体以活动页面为准。

English

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 Beta

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 →

FigureForge walkthrough: describe paper, retrieve references, generate bitmap draft

▶ Full demo video, 1:58 with audio

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.
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.
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.
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.
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.

Add a paper or figure

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.

中文

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 Beta

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 →

FigureForge 流程演示：输入论文、检索参考图、生成位图初稿

▶ 完整演示视频，1 分 58 秒，有声音

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」。
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。
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。
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。
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。

补充论文或图片

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。

📊 收录规模 / Coverage

Venue 会议	2023	2024	2025	2026	Total 合计
🟣 ICLR	108	135	274	228	745
🟢 ICML	63	131	216	338	748
🔴 NeurIPS	236	363	367	—	966
🟠 CVPR	65	64	96	83	308
🔵 ACL	84	74	168	58	384
🟡 AAAI	59	66	139	37	301
				Total 总计	3452

🏅 高等级论文 / High-tier papers

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

Venue 会议	Best / Outstanding	Oral	Spotlight	Total
🟣 ICLR	10	337	155	502
🟢 ICML	13	247	558	818
🔴 NeurIPS	3	32	237	272

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。

🧭 视觉模式 / Visual patterns

标签描述图的视觉功能，不是论文领域。

Tag 标签	含义 Meaning	代表 Examples
`conceptual`	用视觉隐喻讲清核心概念 / visual metaphor for a core idea	DPO, Tree of Thoughts, HippoRAG
`framework`	系统或智能体模块协作总览 / system or agent overview	HuggingGPT, SWE-agent, MetaGPT
`pipeline`	端到端阶段式数据流 / staged end-to-end data flow	LRM, VidMan, LiveStar
`architecture`	模型内部层与张量连接 / internal layers and tensor connections	BLIP-2, Caduceus, Medusa
`taxonomy`	任务、能力或数据全景 / benchmark overview	EmbodiedBench, SimWorld
`teaser`	经过版式设计的图文混排主视觉 / designed mixed-layout visual	VIMA, PixArt-α

🚀 快速开始 / Quick start

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.

🗃️ 仓库结构 / Structure

├── 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

🔧 数据管线与复现 / Reproduce the pipeline

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

🗺️ Roadmap

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）。

🤝 贡献 / Contributing

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）。

🌟 贡献者 / Contributors

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：

贡献者	贡献内容	PR
@timelic	画廊灯箱共享元素过渡动画（View Transitions），以及作者名单、查看图片 / 论文链接、等级角标、圆角与分隔符等细节打磨	#2 · #3
@TansyZenix	灯箱 Tab 焦点限制、生成 SVG 预览的 `sandbox` 隔离、参考图卡片键盘选择与 ARIA 状态、数据校验 CI（`scripts/validate_gallery.py` + workflow）	#5 · #6 · #7 · #8

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.

⚠️ 版权 / Copyright

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。

📌 引用 / Citation

@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}
}

🖼️ 精选图录 / Curated catalog

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…}