Turn AI papers into polished conference slides — with LaTeX-only visuals, default PPTX + English DOCX delivery, and clean AI conference style.

Overview · Quick Start · Features · Citation · Star History

ai-paper2slide-skill is an open-source ChatGPT Skill for converting AI and scientific papers into conference-quality slide packages.

It is designed for papers submitted to or presented at top-tier venues such as NeurIPS, ICML, ICLR, CVPR, ACL, KDD, WWW, AAAI, SIGIR, EMNLP, ECCV, ICCV, MICCAI, and related AI, machine learning, and computer vision conferences.

Unlike generic document-to-slide tools, this Skill focuses on two strict production guarantees:

1. LaTeX-only slide images
   Any image inserted into the slide deck must come from the user-provided LaTeX source package. The Skill does not use PDF screenshots, web images, generated images, previous-conversation images, or external substitutes.

2. Default two-file delivery
   Every full paper-to-slide run returns at least two user-facing files by default: a .pptx slide deck and a .docx English per-slide speaker script.

Upload the LaTeX source package, and I will generate a conference-style PPTX deck plus an English DOCX speaker script.

1. Original LaTeX source package
   Supported formats: .zip, .tar, .tar.gz

2. Compiled paper PDF
   Used for text cross-checking, section order, and final paper layout reference.

• paper2slide_deck.pptx
• paper2slide_speaker_script.docx
• visual_source_manifest.json
• paper2slide_quality_report.md

The Skill prioritizes the original LaTeX source package and scans it for:

• \includegraphics paths
• figure and table environments
• captions
• labels
• surrounding explanatory text
• architecture and method diagrams
• experimental result tables
• ablation and qualitative visualizations

Only visual assets found inside the user-provided LaTeX package are eligible for slide images.

For every full paper-to-slide request, the Skill produces by default:

The Skill guides ChatGPT to create a clean, presentation-ready .pptx deck with:

• 16:9 widescreen layout
• claim-based slide titles
• sparse and readable text
• source-grounded architecture and result visuals
• strong method storytelling
• clear experiment and ablation slides
• final takeaway and impact slide

The default speaker script is English. Users may still request different language modes for slide text, notes, or additional scripts:

Convert this LaTeX paper package into a 10-minute NeurIPS-style presentation.

Please create:
1. A 16:9 PowerPoint slide deck.
2. A Word document with the English per-slide speaking script.
3. A visual source manifest for all architecture figures and experimental result tables.
4. A quality report checking LaTeX-only image compliance, figure/table placement, and script timing.

Only use images that are included in the provided LaTeX package.

请将这篇论文转换为一个 10 分钟中文组会汇报。

请生成：
1. 中文 PPT。
2. 默认英文逐页演讲稿 Word 文档。
3. 可选中文逐页演讲稿 Word 文档。
4. 图表来源 manifest 和质量报告。

Slide 中的所有图片必须只来自我上传的 LaTeX 源文件包。

Scans a LaTeX project or archive and creates a visual source manifest.

python scripts/inspect_latex_assets.py paper_source.zip \
  --output visual_source_manifest.json

Checks whether a slide visual map only references valid figure and table IDs from the manifest.

python scripts/validate_visual_sources.py \
  visual_source_manifest.json \
  slide_visual_map.json \
  --strict-latex-images \
  --output paper2slide_quality_report.md

ai-paper2slide-skill/
├── README.md
├── docs/
│   └── i18n/
│       └── README_ZH.md
├── LICENSE
├── assets/
│   └── README.md
├── references/
│   ├── ai_conference_style_guide.md
│   ├── latex_visual_localization.md
│   ├── quality_checklist.md
│   └── speaker_script_guide.md
└── scripts/
    ├── inspect_latex_assets.py
    └── validate_visual_sources.py

The Skill verifies where every slide image comes from before designing the slide around it.

If an image is not in the user-provided LaTeX source package, it does not go into the deck.

A complete run should always return the PowerPoint deck and the English per-slide speaker script.

Each slide should communicate one research claim, supported by one source-grounded visual or concise evidence.

The deck should look like a polished AI conference presentation, not a generic corporate template.

• The Skill works best when the LaTeX source package is complete and well-organized.
• PDF-only workflows cannot provide paper images under the default strict image policy.
• Complex TikZ figures, rasterized tables, or missing assets may require manual checking.
• The Skill provides a workflow and helper scripts; final artifact generation depends on the ChatGPT environment and available document/slide tools.

Potential future improvements include:

• stricter LaTeX asset provenance checking
• better table-to-slide summarization
• built-in PPTX template themes
• venue-specific slide styles
• automatic timing calibration from generated scripts
• poster-to-slide conversion
• multilingual supplemental speaker scripts
• bilingual slide templates
• integration with open-source presentation agents

Contributions are welcome.

Useful contribution areas include:

• better LaTeX parsing
• improved slide quality checks
• additional conference style guides
• PPTX template assets that do not include external images
• evaluation examples
• multilingual documentation
• support for more paper formats

Please open an issue or pull request with a clear description of the proposed improvement.

If this Skill helps your research presentation workflow, please consider citing or linking to the repository.

@misc{ai-paper2slide-skill,
  title        = {AI Paper2Slide Skill: Conference-Grade Paper-to-Slide Generation for AI Research},
  author       = {Zhixiang Lu},
  year         = {2026},
  howpublished = {\url{https://github.com/Leo1998-Lu/ai-paper2slide-skill}},
  note         = {Open-source ChatGPT Skill for LaTeX-only, conference-style paper-to-slide generation.}
}