PaperView (ppv)
Turn a research paper or codebase into a narrated, animated explainer video, rendered locally with code, not generated.
Here's a video explaining PaperView, generated by PaperView:
https://github.com/user-attachments/assets/9c2df835-8cea-4220-88e0-ae31b8c58a19
Status: v0.1 (early, first public release). Claude Code adapter only.
Demos
Every video below was generated by PaperView itself, explaining real papers and codebases:
| Source | Type | Watch |
|---|
| nanoGPT — GPT-2 in ~600 lines of Python | codebase | ▶ YouTube |
| Attention Is All You Need — the Transformer | paper | ▶ YouTube |
| PagedAttention / vLLM — paging the KV cache | paper | ▶ YouTube |
| SentrySearch — semantic search over video | codebase | ▶ YouTube |
| llama.cpp — ggml, GGUF & quantization | codebase | ▶ YouTube |
| Diffusion models — how AI turns noise into pictures, explained simply | paper · ELI5 | ▶ YouTube |
Contents
- How it works
- Getting Started
- How is this different from NotebookLM?
- CLI
- License
How it works
PaperView splits the job between the coding agent (the judgment) and a deterministic CLI (the mechanics):
- You point the agent at a source — a paper (a local PDF, or an arXiv id/link that's downloaded for you) or a codebase (a local path, or a GitHub URL that's cloned for you).
- The agent reads it and writes a scene plan — the narration plus which visual component (and its props) for each scene, grounded in the paper's text and figures or the code itself.
ppv parses the source — text and figures via PyMuPDF, with real equations pulled from the arXiv LaTeX.ppv tts synthesizes the narration locally with Kokoro — no API keys, nothing leaves your machine.ppv render renders the scene plan deterministically with Remotion, typesetting math with KaTeX — real code, figures, and diagrams, not generated pixels.- The output is a topic-named MP4 in its own run directory under
~/.paperview/runs/.
Getting Started
Prerequisites: Claude Code, plus git, python3, node/npm, and ffmpeg on your PATH. /ppv:setup installs everything else (Python venv, Remotion deps, local models).
Run these inside an interactive Claude Code session.
1. Add the marketplace and install the plugin (one time):
/plugin marketplace add ssrajadh/paperview
/plugin install ppv@paperview
/reload-plugins
2. Bootstrap the toolchain (one time):
/ppv:setup
3. Generate a video — from a paper (a local PDF):
/ppv:gen ~/Downloads/attention_is_all_you_need.pdf, 3 minutes, focus on attention
…or from an arXiv id / link (the PDF is downloaded for you, equations pulled from the LaTeX):
/ppv:gen arxiv.org/abs/1706.03762, 3 minutes, expert
…or from a codebase / GitHub URL (the repo is cloned for you):
/ppv:gen https://github.com/ssrajadh/sentrysearch, 2 minutes, focus on the architecture
To update later, refresh the marketplace and reload (re-run /ppv:setup if a release changes the toolchain):
/plugin marketplace update paperview
/reload-plugins
git clone https://github.com/ssrajadh/paperview && cd paperview
claude --plugin-dir ./adapters/claude-code # load the plugin from the working tree
/ppv:setup
/ppv:gen ~/Downloads/attention_is_all_you_need.pdf, 3 minutes, focus on attention
How is this different from NotebookLM?
NotebookLM's Video Overviews are great for a quick narrated summary. PaperView optimizes for something different — a faithful, reproducible explainer built from the source's real content:
| PaperView | NotebookLM Video Overviews |
|---|
| Visuals | The source's own figures, real typeset equations (KaTeX), syntax-highlighted real code, and code-drawn diagrams | Narrated, mostly generated slides |
| Editable & reproducible | The agent emits a scene plan (plain JSON) you can inspect and tweak; rendering that plan is deterministic — same plan, same video | One-shot cloud generation, not exposed or editable |
| Where it runs | Render and narration are local (Remotion + Kokoro); no upload to a video-generation cloud | Cloud service — sources are processed by Google |
| Codebases | First-class — reads the real source and pastes actual snippets | Built around documents |
| Control | Steer length, focus, depth, and voice; the scene plan is editable JSON; fully open source | Prompt-level steering |
In short: reach for NotebookLM when you want a fast overview; reach for PaperView when you want the paper's actual equations and figures — or a codebase's actual code — in a video you can reproduce, tweak, and run offline.
CLI (what the agent drives)
ppv doctor # health-check the toolchain
ppv parse <pdf|md|txt> --out <dir> # text + figures (PDF, Markdown, or text source)
ppv validate <scenes.json> [--assets <dir>] # fast-fail plan check (no TTS/render cost)
ppv tts <scenes.json> --out <dir> # narration WAVs + durations.json
ppv render <scenes.json> --workdir <dir> --out <mp4> [--resolution 810p|540p] [--draft] [--crf N] [--progress]
Run ppv components for the component library and ppv schema for the full plan JSON Schema
(ppv tts --list-voices lists the narration voices).
License
PaperView itself is licensed under Apache-2.0 (see LICENSE).
It builds on third-party tools that carry their own licenses, which govern what you may do
with the rendered pipeline — read them before any commercial or at-scale use:
- Remotion (rendering) is source-available, not classic OSS: free for individuals and small
teams, but companies above its size threshold need a paid license, and there are restrictions on
building competing products. See https://remotion.dev/license — PaperView's Apache-2.0 grant does
not override these terms.
- PyMuPDF (PDF parsing) is AGPL-3.0 (or a commercial license from Artifex). You install it as
a dependency; obligations attach to anyone who redistributes a combined work.
- Kokoro (local TTS) ships Apache-2.0 weights.
Privacy: PaperView runs entirely locally and collects no data — no analytics, telemetry, or
phone-home. See PRIVACY.md.
