cvpr-2026-poster-skill

内容来源：README.md（说明文档） · 原始地址 · 查看安装指南

原始内容

# CVPR 2026 Poster Skill

Turn a CVPR 2026 paper into a print-ready 84"×42" poster by chatting
with an AI agent. No manual templating, no PowerPoint, no fighting
with print dialogs.

This is a Claude Code / Codex **skill** — not a standalone CLI tool.
You install it once into your agent, then talk to it in plain language.

---

## How to start

### 1. Install the skill (one time)

**Claude Code**

```bash
git clone https://github.com/yunyiliu/cvpr-2026-poster-skill.git
mkdir -p ~/.claude/skills
ln -s "$(pwd)/cvpr-2026-poster-skill/cvpr-2026-poster" ~/.claude/skills/cvpr-2026-poster
```

Start a new Claude Code session (existing sessions don't auto-pick-up
new skills). The skill appears in the available-skills list as
`cvpr-2026-poster`.

**Codex / other agents**: copy or symlink the same `cvpr-2026-poster/`
folder into your agent's skill location — refer to that tool's docs
for the exact path.

### 2. Ask the agent to build the poster

In a new session, write a single message like:

> Use `$cvpr-2026-poster` to create a poster workspace from my Overleaf
> folder at `/path/to/your/overleaf-folder/`. Fill the brief from
> LaTeX, copy figures, auto-fetch institution logos, and sync the
> editable poster.

The agent will:

- Run `init_poster_project.py` to scaffold the workspace
- Run `fill_brief_from_latex.py` to extract your paper's title,
  authors, affiliations, abstract, and figures
- Run `sync_poster_from_brief.py` to populate the editable HTML, copy
  figures into place, and fetch logos from your institutions' websites
- Open the resulting `poster/index.html` for you to inspect

If you have multiple institutions, list them in the same order as
`Affiliations` in `poster_brief.md` and add one institution website
per school. The agent does this automatically when the LaTeX is clear.

### 3. Iterate by talking to the agent

Once the poster is open in the browser, just describe edits in chat:

> Make the Results card use a side-by-side figure/text layout, and
> shrink the conference logo a bit.

> The Contributions card is too long — keep it to one lead sentence
> and three bullets.

> Move the Conclusion card to the bottom of the rightmost column.

The agent will edit `poster-config.json` (and re-embed it into
`index.html`) for you. Hard-refresh the browser tab to see the result.

You can also edit in the browser directly:

- Drag column dividers to change column widths
- Drag horizontal dividers to change card heights
- Drag a card's `◆` handle (or click two handles in sequence) to
  swap card positions
- Use `A+` / `A-` to scale all fonts globally
- Click **Save** when you're happy — that downloads
  `poster-config.json` to `~/Downloads/` with your layout

### 4. Export to a print-ready PDF

After you click **Save** in the browser, in a terminal:

```bash
cd poster-workspace
bash bake_and_export.sh
```

This picks up the saved `poster-config.json` from `~/Downloads/`,
re-embeds it into `index.html`, and runs headless Chrome to produce
`poster.pdf` at exactly **84.01" × 42.01"** (or the workshop size).

If you have not touched the layout in the browser since the last
agent edit, just run:

```bash
bash export_pdf.sh
```

Send `poster.pdf` directly to your printer.

---

## What you need

- Google Chrome (for the PDF export step)
- Python 3 (preinstalled on macOS and most Linux distros)

That's all. The skill ships the rest:

- LaTeX extractor
- Institution-logo auto-fetcher
- Editable HTML poster scaffold
- Headless-Chrome PDF exporter at the exact poster size
- Official CVPR 2026 Denver logo and template

---

## What gets created

A typical workspace looks like:

```
poster-workspace/
├── poster_brief.md         ← human-readable summary of paper content
├── poster/
│   ├── index.html          ← the editable poster, open in a browser
│   ├── poster-config.json  ← serialized layout + content
│   ├── figures/            ← display copies of figures
│   └── logos/
│       ├── user/           ← school logos (top-left)
│       └── official/       ← CVPR conference logo (top-right)
├── assets/
│   ├── figures/            ← original figures from the LaTeX project
│   └── logos/              ← school / lab logos (user or auto-fetched)
├── references/
│   └── latex-extract.md    ← summary of what was pulled from LaTeX
├── export_pdf.sh           ← one-command headless-Chrome PDF export
└── bake_and_export.sh      ← bake current browser layout, then export
```

---

## CVPR 2026 size reference

| Track | Poster size | Aspect ratio |
|-------|-------------|--------------|
| Main / Findings | 84" × 42" (2134 × 1067 mm) | 2 : 1 landscape |
| Workshop | 42" × 21" (1067 × 533 mm) | 2 : 1 landscape |

Output is PDF, no bleed. The exporter honors the poster's `@page`
declaration so the PDF page size matches the track exactly.

---

## Without an AI agent

If you want to drive the scripts manually instead of through an agent,
all of them are plain Python:

```bash
python3 cvpr-2026-poster/scripts/init_poster_project.py --project-dir ./poster-workspace --track main
python3 cvpr-2026-poster/scripts/fill_brief_from_latex.py --project-dir ./poster-workspace --latex-dir /path/to/overleaf --copy-figures
# edit poster-workspace/poster_brief.md
python3 cvpr-2026-poster/scripts/sync_poster_from_brief.py --project-dir ./poster-workspace --fetch-logos-if-missing
open poster-workspace/poster/index.html
# (edit in browser, click Save when done)
cd poster-workspace && bash bake_and_export.sh
```

---

## License

MIT. See [LICENSE](https://github.com/yunyiliu/cvpr-2026-poster-skill/tree/HEAD/LICENSE).

CVPR 2026 Poster Skill

Turn a CVPR 2026 paper into a print-ready 84"×42" poster by chatting with an AI agent. No manual templating, no PowerPoint, no fighting with print dialogs.

This is a Claude Code / Codex skill — not a standalone CLI tool. You install it once into your agent, then talk to it in plain language.

How to start

1. Install the skill (one time)

Claude Code

git clone https://github.com/yunyiliu/cvpr-2026-poster-skill.git
mkdir -p ~/.claude/skills
ln -s "$(pwd)/cvpr-2026-poster-skill/cvpr-2026-poster" ~/.claude/skills/cvpr-2026-poster

Start a new Claude Code session (existing sessions don't auto-pick-up new skills). The skill appears in the available-skills list as cvpr-2026-poster.

Codex / other agents: copy or symlink the same cvpr-2026-poster/ folder into your agent's skill location — refer to that tool's docs for the exact path.

2. Ask the agent to build the poster

In a new session, write a single message like:

Use $cvpr-2026-poster to create a poster workspace from my Overleaf folder at /path/to/your/overleaf-folder/. Fill the brief from LaTeX, copy figures, auto-fetch institution logos, and sync the editable poster.

The agent will:

Run init_poster_project.py to scaffold the workspace
Run fill_brief_from_latex.py to extract your paper's title, authors, affiliations, abstract, and figures
Run sync_poster_from_brief.py to populate the editable HTML, copy figures into place, and fetch logos from your institutions' websites
Open the resulting poster/index.html for you to inspect

If you have multiple institutions, list them in the same order as Affiliations in poster_brief.md and add one institution website per school. The agent does this automatically when the LaTeX is clear.

3. Iterate by talking to the agent

Once the poster is open in the browser, just describe edits in chat:

Make the Results card use a side-by-side figure/text layout, and shrink the conference logo a bit.

The Contributions card is too long — keep it to one lead sentence and three bullets.

Move the Conclusion card to the bottom of the rightmost column.

The agent will edit poster-config.json (and re-embed it into index.html) for you. Hard-refresh the browser tab to see the result.

You can also edit in the browser directly:

Drag column dividers to change column widths
Drag horizontal dividers to change card heights
Drag a card's ◆ handle (or click two handles in sequence) to swap card positions
Use A+ / A- to scale all fonts globally
Click Save when you're happy — that downloads poster-config.json to ~/Downloads/ with your layout

4. Export to a print-ready PDF

After you click Save in the browser, in a terminal:

cd poster-workspace
bash bake_and_export.sh

This picks up the saved poster-config.json from ~/Downloads/, re-embeds it into index.html, and runs headless Chrome to produce poster.pdf at exactly 84.01" × 42.01" (or the workshop size).

If you have not touched the layout in the browser since the last agent edit, just run:

bash export_pdf.sh

Send poster.pdf directly to your printer.

What you need

Google Chrome (for the PDF export step)
Python 3 (preinstalled on macOS and most Linux distros)

That's all. The skill ships the rest:

LaTeX extractor
Institution-logo auto-fetcher
Editable HTML poster scaffold
Headless-Chrome PDF exporter at the exact poster size
Official CVPR 2026 Denver logo and template

What gets created

A typical workspace looks like:

poster-workspace/
├── poster_brief.md         ← human-readable summary of paper content
├── poster/
│   ├── index.html          ← the editable poster, open in a browser
│   ├── poster-config.json  ← serialized layout + content
│   ├── figures/            ← display copies of figures
│   └── logos/
│       ├── user/           ← school logos (top-left)
│       └── official/       ← CVPR conference logo (top-right)
├── assets/
│   ├── figures/            ← original figures from the LaTeX project
│   └── logos/              ← school / lab logos (user or auto-fetched)
├── references/
│   └── latex-extract.md    ← summary of what was pulled from LaTeX
├── export_pdf.sh           ← one-command headless-Chrome PDF export
└── bake_and_export.sh      ← bake current browser layout, then export

CVPR 2026 size reference

Track	Poster size	Aspect ratio
Main / Findings	84" × 42" (2134 × 1067 mm)	2 : 1 landscape
Workshop	42" × 21" (1067 × 533 mm)	2 : 1 landscape

Output is PDF, no bleed. The exporter honors the poster's @page declaration so the PDF page size matches the track exactly.

Without an AI agent

If you want to drive the scripts manually instead of through an agent, all of them are plain Python:

python3 cvpr-2026-poster/scripts/init_poster_project.py --project-dir ./poster-workspace --track main
python3 cvpr-2026-poster/scripts/fill_brief_from_latex.py --project-dir ./poster-workspace --latex-dir /path/to/overleaf --copy-figures
# edit poster-workspace/poster_brief.md
python3 cvpr-2026-poster/scripts/sync_poster_from_brief.py --project-dir ./poster-workspace --fetch-logos-if-missing
open poster-workspace/poster/index.html
# (edit in browser, click Save when done)
cd poster-workspace && bash bake_and_export.sh

License

MIT. See LICENSE.