| name | poster-video |
| description | Use when the user wants to turn a set of slide / poster images into a CVPR-oral-style narrated video. The skill scaffolds a workspace, lets the agent write per-slide narration into script.md, synthesizes voice with macOS `say` or OpenAI TTS, and assembles the final MP4 with ffmpeg. Pairs naturally with the cvpr-2026-poster skill but accepts any folder of slide images. |
Poster / Slide → Video
Overview
This skill turns an ordered set of slide images plus a per-slide narration
script into a single MP4 with voiceover, suitable for a 5–10 minute CVPR
oral, a workshop spotlight, or a recap clip. It is not tied to any
specific source format — pass in PNG/JPG images, get a video.
It pairs well with the cvpr-2026-poster skill: export the editable
poster to PDF, rasterize selected regions as slide images, then run this
skill to narrate and assemble.
Workflow
- Scaffold a
video-workspace/ with init_video_project.py.
- Get slide images into
video-workspace/slides/ by either:
- dropping in your own PNG/JPG files named in order
(
01-title.png, 02-problem.png, ...), or
- running
import_deck.py --input deck.pptx --slides-dir slides/
to auto-convert a .pptx, .key, or .pdf into per-page PNGs.
- Fill or generate
script.md — one ## NN section per slide.
- Run
synthesize_voice.py to convert each script section to audio.
- Run
render_video.py to assemble output/poster_video.mp4.
Deck input formats
import_deck.py accepts:
.pptx — converted via LibreOffice (brew install --cask libreoffice)
or macOS Keynote (free, preinstalled)..key — converted via Keynote..pdf — rasterized page-by-page; works for slide-deck PDFs as-is,
and for paper PDFs if you want a "scrolling through the paper" video.
Each input page becomes one slide image named 01.png, 02.png, ...
in the workspace's slides/ folder.
Requirements
- Python 3 (preinstalled on macOS / most Linux)
- ffmpeg (
brew install ffmpeg)
- For TTS: macOS
say (default, free, no key) OR OPENAI_API_KEY for
high-quality OpenAI TTS
Typical prompts
Use $poster-video to turn the slides in ~/Desktop/my-slides into a 5-minute CVPR oral video.Use $poster-video to draft a per-slide narration for these images, matching a confident-but-clear CVPR-oral tone.Use $poster-video to re-render the video with the OpenAI TTS voice alloy and 1.05x speed.
Default video format
- Resolution: 1920×1080 (16:9)
- Audio: AAC 192 kbps, 44.1 kHz
- Video codec: H.264, yuv420p, ~24 fps
- Each slide held for the duration of its narration; cross-fades
optional via
--crossfade-ms
Voice options
| Provider | When to use | Setup |
|---|
macOS say | Free; quick draft / preview | None (built-in) |
| OpenAI TTS | Conference-grade voice for the final cut | Set OPENAI_API_KEY |
Script structure
script.md uses one second-level heading per slide. The number must
match the image filename prefix:
## 01
Welcome to our CVPR 2026 talk on Your Paper Title. Today we ask the
question that motivates this work, in one sentence the audience can
remember.
## 02
Cross-entropy training raises the probability of the reference token
but never suppresses the wrong token the model actually chose. ...
Tone guidance the agent should respect when writing the script:
- 90–120 seconds per slide for an oral talk; 20–40 seconds for a recap
- One sentence per breath group — TTS reads punctuation literally
- Spell out symbols (
p hat t) — TTS mangles math notation
- Avoid em-dashes inside numbers (
0.143, not 0—143)
Output
video-workspace/
├── slides/ # input images
├── script.md # narration, one ## NN per slide
├── audio/ # one .wav or .mp3 per slide, generated
├── output/
│ └── poster_video.mp4 # final video
└── README-workspace.md
