AutoYouTube

Automated pipeline that turns Hugging Face Daily Papers (or a single PDF) into:

• slide deck images,
• narrated audio tracks,
• subtitle-burned MP4 videos,
• optional YouTube uploads.

The repository is built for daily research-content production with minimal manual steps.

Table of Contents

• Overview
• Pipeline
• Tech Stack
• Repository Structure
• Requirements
• Quick Start
• Usage
• CLI Reference
• Environment Variables
• Outputs
• YouTube Upload Behavior
• Figure Extraction (Optional but Built-in)
• Troubleshooting
• Publishing Checklist
• Known Limitations
• License

Overview

main.py orchestrates a full content pipeline:

1. Fetch top papers from Hugging Face Daily Papers (or use --pdf-url).
2. Resolve PDF links and download PDFs.
3. Extract core paper text (abstract, introduction, conclusion, plus full text).
4. Ask an LLM for structured slide content in JSON.
5. Build Marp markdown with modern CSS and optional figure embedding.
6. Render slides to PNG.
7. Generate TTS audio per slide and per language.
8. Compose video with subtitles using MoviePy.
9. Upload to YouTube (optional).

Pipeline

Hugging Face Daily Papers / --pdf-url
  -> PDF download
  -> PDF text + figure extraction
  -> LLM summarization -> slide specs
  -> Marp markdown generation
  -> Marp PNG rendering
  -> OpenAI TTS (per slide, per language)
  -> MoviePy composition + hard subtitles
  -> MP4 output
  -> Optional YouTube upload

Tech Stack

• Python (pipeline orchestration)
• OpenAI API (LLM summarization, translation, TTS)
• PyMuPDF (PDF text extraction)
• Marp CLI (slide rendering)
• MoviePy + FFmpeg (video composition/encoding)
• YouTube Data API v3 (upload)
• Optional: LayoutParser + Detectron2 (figure/table extraction)

Repository Structure

main.py                  # Entry point
config.py                # Env-driven configuration

daily_papers/            # HF fetch + PDF handling + text extraction
llm/                     # LLM client, prompts, summarizer, translator
slides/                  # Figure handling, markdown builder, Marp renderer
tts/                     # OpenAI TTS client
video/                   # MoviePy video builder + subtitles
youtube/                 # YouTube OAuth/upload
storage/                 # Output path helpers
scripts/                 # Setup and convenience scripts
models/publaynet/        # PubLayNet config + weights (for figure extraction)

Requirements

Runtime

• Python 3.10+ recommended
• ffmpeg on PATH
• marp CLI on PATH
• OpenAI API key

Install Python dependencies

pip install -r requirements.txt

Install system dependencies

# macOS example
brew install ffmpeg node
npm i -g @marp-team/marp-cli

Quick Start

1. Create virtual environment

./scripts/setup_venv.sh

2. Configure environment

cp .env.example .env

Set at least:

• OPENAI_API_KEY

If uploading to YouTube, also set:

• YOUTUBE_CLIENT_SECRETS_FILE
• YOUTUBE_TOKEN_FILE

Optional but useful to add in .env:

LANGUAGES=en,ko
TTS_SPEED=1.2
TTS_STYLE_INSTRUCTION=

3. Activate environment

source scripts/env.sh

4. Run pipeline (local video build, no upload)

python main.py --date 2026-02-21 --top-k 10 --skip-upload

Usage

Daily papers mode

python main.py --date 2026-02-21 --top-k 10 --skip-upload

Video-only shortcut (always skips upload)

python main.py --date 2026-02-21 --top-k 10 --video-only
# or
./scripts/run_video_only.sh 2026-02-21 10

Single PDF mode

python main.py \
  --pdf-url https://arxiv.org/pdf/2511.21689.pdf \
  --paper-id 2511.21689 \
  --paper-title "Your Paper Title" \
  --origin "Your Lab/Company" \
  --skip-upload

Partial runs (debug/dev)

python main.py --date 2026-02-21 --skip-render
python main.py --date 2026-02-21 --skip-tts
python main.py --date 2026-02-21 --skip-video

CLI Reference

Flag	Description
--date YYYY-MM-DD	Target date. Defaults to today.
--top-k N	Number of papers to process. Defaults to TOP_K/10.
--languages en,ko,...	Narration languages. First language is primary output video.
--pdf-url URL	Process a single PDF instead of Daily Papers fetch.
--paper-id ID	Optional custom ID in single-PDF mode.
--paper-title TITLE	Optional custom title in single-PDF mode.
--origin TEXT	Optional affiliation/origin override for intro narration.
--skip-render	Skip Marp PNG rendering.
--skip-tts	Skip TTS generation.
--skip-video	Skip video composition.
--skip-upload	Disable YouTube upload.
--video-only	Build local video and skip upload (same effect as --skip-upload for upload phase).

Environment Variables

Variable	Default	Description
HF_BASE_URL	https://huggingface.co	Hugging Face base URL.
TOP_K	10	Default top-K paper count.
OPENAI_API_KEY	-	Required for LLM/TTS steps.
OPENAI_LLM_MODEL	gpt-5.3	Model used for summarization and translation.
OPENAI_TTS_MODEL	gpt-4o-mini-tts	TTS model.
OPENAI_TTS_VOICE	alloy	Voice preset.
TTS_STYLE_INSTRUCTION	empty	Style hint string (stored, currently not injected into spoken text).
TTS_SPEED	1.2	TTS speed multiplier, clamped to [0.5, 4.0].
LANGUAGES	en,ko	Comma-separated target narration languages.
YOUTUBE_CLIENT_SECRETS_FILE	empty	OAuth client JSON path.
YOUTUBE_TOKEN_FILE	empty	OAuth token cache path.
OUTPUT_BASE_DIR	./outputs	Base output directory.
LOG_LEVEL	INFO	Python logging level.

Outputs

Daily papers mode (--date)

outputs/{date}/
  daily_papers_{date}.mp4
  daily_papers_{date}_{lang}.mp4
  slides/
    slides_{date}.md
    slides_{date}_*.png
    scripts_{date}_{lang}.txt
  {paper_id}/
    paper.pdf
    captions.json
    figures/*.png
  audio/{lang}/
    audio_slide_001.mp3
    ...

Single PDF mode (--pdf-url)

outputs/{date}/{paper_id}/
  paper.pdf
  captions.json
  figures/*.png
  slides/
    slides_{date}.md
    slides_{date}_*.png
    figures/*.png          # copied assets used in slide markdown
  scripts_{date}_{lang}.txt
  audio/{lang}/audio_slide_*.mp3
  daily_papers_{date}.mp4
  daily_papers_{date}_{lang}.mp4

YouTube Upload Behavior

• Upload occurs only when both conditions are met:
  • you did not pass --skip-upload / --video-only
  • both YOUTUBE_CLIENT_SECRETS_FILE and YOUTUBE_TOKEN_FILE are configured
• First OAuth run opens a local browser for consent.
• Videos are uploaded as:
  • privacyStatus: unlisted
  • categoryId: 28 (Science & Technology)
• In multi-language output, title suffix includes language labels (for non-primary tracks).

Figure Extraction (Optional but Built-in)

Figure extraction runs per downloaded PDF before summarization.

Behavior:

• If captions.json already exists, extraction is skipped (cache behavior).
• If model files or optional deps are missing, extraction is skipped gracefully.
• Extracted figure/table metadata is fed into the LLM prompt to improve slide quality.

Model/env paths:

• Default model files:
  • models/publaynet/config.yaml
  • models/publaynet/model_final.pth
• Override via:
  • PUBLayNET_CONFIG
  • PUBLayNET_WEIGHTS

Optional dependencies for this feature are not in requirements.txt by default (e.g., layoutparser, detectron2, opencv-python, numpy).

Troubleshooting

marp CLI not found

Install Marp CLI globally and verify:

npm i -g @marp-team/marp-cli
marp --version

FFmpeg / video encoding errors

Install and verify:

ffmpeg -version

OpenAI request failures

Check:

• OPENAI_API_KEY is set correctly
• model names are valid for your account
• quota/rate limits are not exhausted

No video built

Common causes:

• slide image count and audio file count mismatch
• --skip-render or --skip-tts used in ways that leave missing artifacts
• upstream LLM/TTS failures in logs

YouTube upload errors

Check:

• OAuth client JSON path is valid
• token path is writable
• if token is stale, delete token JSON and authenticate again

Publishing Checklist

Before pushing to GitHub:

• Remove or ignore secrets:
  • .env
  • OAuth token files
  • OAuth client secrets
• Avoid committing large generated artifacts:
  • outputs/
  • local MP4/PNG/MP3 files
• Add a proper .gitignore if not present.
• Keep LICENSE and NOTICE files when redistributing.

Known Limitations

• No automated test suite is included yet.
• LLM output quality depends heavily on model behavior and prompt adherence.
• TTS is generated slide-by-slide; long runs can be costly/time-consuming.
• Figure extraction quality depends on external model/deps and PDF layout quality.

License

This project is licensed under the Apache License 2.0. See LICENSE.

If you redistribute this project or derivative works, you should:

• include the LICENSE file,
• retain the NOTICE file,
• keep attribution to the original project/author in source or documentation.