English | 简体中文
Group meetings, conference talks, thesis proposals, project updates. Point it at a quarterly report instead and it won't complain, but slides for research work are what it was built around.
The document on the left of that picture is docs/demo/gold-price-2024.pdf, shipped with the repo so you can run the same thing yourself.
You upload a paper — or several PDFs that read in order as one pack — or just paste text, pick a template, and a vision model reads the pages and writes a complete main.tex against that template. XeLaTeX compiles it. About thirty to sixty seconds later you have a PDF, the .tex behind it, and a zip you can drop straight into Overleaf. Slides come back in English whatever language the source is in; ask for another language in the instructions and you get that instead.
| Key | Name | Origin | License |
|---|---|---|---|
academic |
Academic (Palatino) | collegeBeamer derivative | CC BY 4.0 |
emerald |
Emerald | collegeBeamer derivative | CC BY 4.0 |
graphite |
Graphite | collegeBeamer derivative | CC BY 4.0 |
navy |
Navy | collegeBeamer derivative | CC BY 4.0 |
scholar |
Scholar | collegeBeamer derivative | CC BY 4.0 |
simpleplus |
SimplePlus | pm25/SimplePlus-BeamerTheme | Unlicense |
execushares |
Execushares | hamaluik/Beamer-Theme-Execushares | MIT |
The collegeBeamer packs share an ancestor but are distinct families now: Academic is a university-press book plate (centered small caps between double hairlines, roman-numeral dividers), Emerald is editorial (display serif over flowing curves, full-bleed part pages with ghost numerals), Graphite is Swiss typographic (sans-first Heros, hard grid, monochrome charts), Navy is a boardroom deck (banded white-on-navy headers, split-panel agendas, a footer progress line), and Scholar keeps the line's original composition — a full-page engraved plate beside the title block. SimplePlus and Execushares keep their own upstream identities. All seven compile with XeLaTeX on a plain TeX Live install with no extra fonts registered — the faces each theme needs ship inside its pack or resolve from TeX Live by name. Chinese input works in every pack. None of them ship anyone's logo — the images under src/brand/ are neutral placeholders meant to be replaced.
This is the one idea the project is built on, and it is worth stating plainly because it explains most of the code that isn't here.
Every template directory contains a main.tex that is a complete, compilable presentation: a title page, a table of contents, worked examples of each layout, real pgfplots charts, and a comment block at the top addressed to the model. That file is handed to the model verbatim. There is no rule sheet next to it, no style DSL, no prompt library. The model is shown a deck that looks right and asked to write another one.
This has a consequence worth knowing before you edit a template. When the layout examples contained an empty rectangle captioned "place a chart here", every generated deck came back with an empty rectangle captioned with the topic. When one layout put text where a figure belonged, the model invented an \includegraphics pointing at a file that did not exist. Neither was solved by adding instructions; both were solved by making the example better. The model is not disobeying — it is copying, accurately, whatever it was shown. Put a real chart in your template and you get real charts back.
An earlier version of this idea went the other way. It grew to about 31,000 lines, 82 routes and a five-stage pipeline that summarised the PDF, then summarised the summary, then truncated the result to fit a prompt budget. What actually reached the model was around 2.2 KB of text per document. It scored well on its own metrics and produced decks nobody wanted. TeXDeck is the reaction: keep the handoff to the model honest, and keep the code small enough that the handoff is still the whole product.
Two smaller rules follow from the same instinct. Source pages go to the model as images, with no intermediate summary and no silent truncation. And when something fails — the model returns unusable TeX, or XeLaTeX still fails after its repair attempts — the job fails with the real error. There is no fallback deck.
The quickest way is Docker — nothing to install but a key:
git clone https://github.com/HuanPP/TeXDeck.git && cd TeXDeck
cat > .env <<'ENV'
TEXDECK_LLM_API_KEY=your-key-here
# Any OpenAI-compatible vision endpoint works; these are the defaults:
TEXDECK_LLM_BASE_URL=https://generativelanguage.googleapis.com/v1beta/openai/
TEXDECK_LLM_MODEL=gemini-pro-latest
ENV
docker compose up --buildThe first build downloads TeX Live, so it takes a while; after that, open http://localhost:8000. Finished decks land in ./artifacts. The image bundles Noto CJK, so Chinese output uses Source Han faces out of the box.
You need Python 3.12+, Node 20+, a TeX Live install with XeLaTeX (synctex too, if you want per-slide editing), and an API key for a vision-capable model.
git clone https://github.com/HuanPP/TeXDeck.git && cd TeXDeck
uv venv .venv --python 3.12
uv pip install -e server --python .venv/bin/python
cp server/.env.example server/.env # then set TEXDECK_LLM_API_KEY
./server/run.sh # API on :8000In a second shell:
cd web && npm install && npm run dev # UI on :5173
The model has to be able to read images — PDF pages are sent as page renders, not as extracted text. Anything with an OpenAI-compatible endpoint works; point TEXDECK_LLM_BASE_URL at it.
Gemini is the recommendation if you have a key. Its multimodal reading is the best of the options here, and it handles a long run of page images without losing the thread:
TEXDECK_LLM_BASE_URL=https://generativelanguage.googleapis.com/v1beta/openai/
TEXDECK_LLM_MODEL=gemini-pro-latest
gemini-pro-latest is an alias that tracks Google's current Pro model, so this keeps working as they ship new ones. Pin a specific version instead (gemini-2.5-pro, gemini-3.1-pro-preview) if you would rather it never change under you.
Moonshot's Kimi works too, and is what this was developed against:
TEXDECK_LLM_BASE_URL=https://api.moonshot.cn/v1
TEXDECK_LLM_MODEL=moonshot-v1-128k-vision-preview
Two things that will not work: text-only models, and the Moonshot kimi-k2.6 / kimi-k3 families, which reject every temperature except 1 and so refuse the calls this makes.
Copy templates/academic to templates/mylab, change the key and name in its template.json, and drop your own logo and background into src/brand/ over the three placeholder files. Restart the server and your template is in the picker.
If you want it to look like something else entirely, edit main.tex — the layouts, the colours, the examples, the comment block at the top telling the model what to do. Whatever you put in that file is what the model will imitate, so it is worth making the examples good. That is the whole extension mechanism; there is nothing else to learn.
Generation can fail. The model occasionally writes LaTeX that will not compile — a table row with one cell too many is the classic — and while there is an automatic repair pass, it does not always land. When that happens you get the error and the compile log rather than a half-broken deck. Running it again usually works — the Regenerate button re-runs the same inputs in one click. Better models fail less.
Editing comes in two sizes. Per-slide edit rewrites one frame from one instruction. Deck-wide refine takes one instruction against the whole deck — "cut it to ten slides", "make the charts bigger", "switch it to Chinese" — through the same validate-compile-repair machinery, and rolls back if the result will not build.
Slide count is a request, not a guarantee. Ask for eight and you will often get ten.
PDFs are capped at 200 pages combined, up to ten files per run. Past 40, the model is shown thumbnails of everything first and asked which pages carry the argument; only those are sent at full resolution.
Per-slide editing maps a PDF page back to its \begin{frame} using synctex. If your TeX installation does not have synctex, that feature tells you so instead of guessing.
cd server && ../.venv/bin/python -m pytest # includes a real XeLaTeX compile of every template
cd server && ../.venv/bin/python -m pytest -m "not slow"The compile suite is the one worth keeping. It renders every bundled template through the real toolchain, which is the only check that catches a template that has quietly stopped working.
MIT for the application code. The bundled themes carry their own licenses and attribution — see NOTICE before redistributing.