---

title: Manhwa Recap Studio emoji: 🎬 colorFrom: green colorTo: purple sdk: docker app_port: 7860 pinned: false

Manhwa Recap Studio

Auto-scrape manhwa/manga/webtoon chapters, transcribe panel text with RapidOCR (PP-OCRv5, ONNXRuntime) as the primary transcriptor (no API keys, no PaddlePaddle framework dependency) with PaddleOCR PP-OCRv4 as a fallback tier (x86 only) and VLM as a last resort, translate to English, and render a narrated recap video with text-to-speech audio.

Self-hosting on a VM (the current production setup)

This is how the project actually runs in production (Oracle Cloud VM, x86 or ARM). One script does everything:

git clone <this-repo-url> manhwa-recap-studio && cd manhwa-recap-studio
SKIP_OLLAMA=1 ./setup.sh          # drop SKIP_OLLAMA on an 8GB+ box if you want local LLMs
# paste secrets into .env (only if you want cloud VLM — the default OCR flow needs none)

setup.sh installs system packages, the Python venv (RapidOCR + optional PaddleOCR), Bun deps, Piper TTS, Caddy, the systemd unit, builds the Next.js bundle, and starts everything. It is idempotent — re-run it any time.

Moving to a fresh instance (incl. ARM): see MIGRATION.md. It covers the aarch64 specifics (PaddleOCR skipped, RapidOCR is primary) and what to carry over from an old box (db/custom.db, data/, .env secrets).

The Docker / Hugging Face Spaces path below still works but is not how the current deployment runs.

Quick Deploy — Hugging Face Spaces (5 steps, ~15 min)

See DEPLOYMENT.md for the full guide. Short version:

1. Create a free Hugging Face account at huggingface.co
2. Create a new Space → SDK: Docker → Space name: manhwa-recap-studio
3. Upload this repo to the Space (or clone + push via git)
4. Enable persistent storage (Settings → Persistent storage → 20 GB free)
5. Wait ~10 min for the Docker build → your app is live at https://yourname-manhwa-recap-studio.hf.space

No environment variables required to start. Add MEGA_EMAIL / MEGA_PASSWORD (optional) for cloud archive to Mega (20 GB free).

What runs inside the container

Service	Port	Purpose
Next.js (standalone)	3000	Frontend + API routes
Pipeline-service	3001	Job queue + VLM fallback + socket.io
PaddleOCR service	3002	RapidOCR PP-OCRv6 (primary) + PaddleOCR PP-OCRv4 (fallback) OCR engine
Caddy	7860	Reverse proxy (HF Spaces entry point)

Optional env vars (Settings → Repository secrets)

Variable	Purpose
MEGA_EMAIL	Mega cloud archive (20 GB free)
MEGA_PASSWORD	Mega cloud archive
GROQ_API_KEY	Better narration/translation (free at console.groq.com/keys)
AUTO_ARCHIVE	true (default) to auto-upload videos to Mega

Features

• Search 6 sources: MangaHere, FanFox, Webtoons, AsuraScans, MAL, AniList
• 55 narration voices with inline preview playback
• PaddleOCR PP-OCRv4 primary transcription (no API keys)
• VLM providers (SiliconFlow, Gemini, Groq) as automatic OCR fallback
• YOLO panel detection
• edge-tts narration with clean audio (no pops/clicks)
• Mega cloud archive (auto-upload + on-demand restore)
• Full HTTP Range video streaming (seek support)

Oracle Cloud CPU-only production mode

Production mode is local-first and does not require edge-tts, OpenAI, or any cloud TTS API. The supported Oracle layout uses an isolated Python 3.10.4 runtime (do not replace /usr/bin/python3) and a dedicated virtual environment:

/opt/python3.10/bin/python3.10 -m venv /opt/manhwa-recap-studio/.venv
source /opt/manhwa-recap-studio/.venv/bin/activate
python --version  # Python 3.10.4
pip install -r pipeline/requirements.txt
pip install -r mini-services/paddleocr-service/requirements.txt
pip install --no-deps rapidocr==3.9.2

Required local binaries for production are ffmpeg, ffprobe, piper, and espeak-ng (or espeak). Set a small CPU-friendly English Piper voice with PIPER_VOICE_MODEL=/opt/piper/voices/en_US-lessac-medium.onnx (or PIPER_VOICE). The production TTS cascade is Piper, Piper retry, then eSpeak NG; non-empty narration is never converted to successful silence.

Production validation commands:

PYTHONPATH=. python pipeline/production_canary.py --work-dir /tmp/mrs-canary --job-id canary
PYTHONPATH=. python pipeline/crash_resume_harness.py --work-dir /tmp/mrs-crash-tts --job-id crash-tts --stage TTS

The SQLite state database is stored under the chosen work directory as state.sqlite; temporary artifacts, promoted artifacts, and quarantined artifacts live under tmp/, artifacts/, and quarantine/ in that work directory.