AI-Powered Video Pipeline

Manhwa Recap
Studio

Enter any manhwa, manga, or webtoon name. We search every source at once, scrape every chapter, transcribe dialogue with AI, and render a narrated recap video.

/focus
·
Escclear
·
Bbookmarks

Recently Viewed

Viewed manga will appear here

How It Works

5-Step Automated Pipeline

From search to finished video in minutes — fully automated

Search

~5s

Enter any manhwa title. We query every source at once — MangaDex, MangaGeko, Webtoons, MangaPill, Comick, MangaHere, FanFox, AsuraScans, WeebCentral, Toonily, plus MAL & AniList.

Download

~30s/ch

Scrapes all panel images from every chapter automatically using source-specific scrapers.

OCR

~15s/ch

PaddleOCR PP-OCRv4 reads speech bubbles and captions from each panel — fast, local, no API keys needed.

Detect Panels

~30s/ch

YOLO + contour-based detection finds panel boundaries. No chopping — each panel stays complete.

Render Video

~2min/ch

Panels are voiced with edge-tts narration and merged into a single recap MP4 with ffmpeg.

Platform Capabilities

Everything you need

A complete toolkit for manhwa recap video creation

Multi-Source Search

Query 6 manga databases simultaneously — MangaHere, FanFox, Webtoons, AsuraScans, MAL & AniList.

6 Sources

Auto Scraping

Source-specific scrapers download every panel image from every chapter, deduplicated across mirrors.

Auto Chapter

PaddleOCR Transcription

PP-OCRv4 reads speech bubbles and captions from each panel — fast, local, no API keys needed.

PP-OCRv4

Panel Detection

YOLO + contour analysis finds panel boundaries ensuring each panel stays complete and un-split.

YOLO AI

Video Rendering

Panels voiced with 55+ TTS voices and merged into a single recap MP4 with ffmpeg.

55+ Voices

Multi-Language

Translate from Korean, Japanese, Chinese, Spanish, French, German and more to English.

9+ Languages

Groq Accelerated

Optional Groq LPU hardware delivers 3-5x faster VLM fallback transcription and narration rewriting.

3-5x Faster

100% Free & Local

Open-source stack with no paid requirements. All AI processing uses free-tier APIs.

$0 Cost
Pipeline Stats

Your Studio Analytics

Overview of your video creation pipeline performance

Job Status Distribution

0 total
Done0
Error0
Cancelled0
Active0
Images Processed

0

Chapters Processed

0

Videos Created

0

Avg Processing Time

—

Success Rate

— (0/0 completed)

Most Used Voice

—

Recent Activity

No recent activity

Job History
FAQ

Frequently Asked Questions

Everything you need to know about the pipeline

The pipeline has 5 stages: (1) Search across 6 sources (MangaHere, FanFox, Webtoons, AsuraScans, MAL, AniList). (2) Download every chapter image. (3) Transcribe speech bubble text using VLM (vision language model). (4) Generate narration with edge-tts text-to-speech. (5) Render the final MP4 video with ffmpeg — panels synced to narration audio. The whole process takes ~6 minutes per chapter.
This happens when all VLM providers are rate-limited (429 errors). The pipeline transcribes panel text using z-ai, Groq, and Gemini — if all three are rate-limited simultaneously, panels get empty text and no narration is generated. The system automatically retries with backoff. Wait 30-60 minutes for the rate limit to reset, then run the job again. The VLM cache also reuses successful transcriptions from previous runs, so re-running a failed job is faster.
VLM transcription is the slowest stage. To speed it up: (1) Set GROQ_API_KEY — Groq's LPU hardware is 3-5x faster than other providers. Get a free key at console.groq.com/keys. (2) Set GEMINI_API_KEY as a second provider. (3) Adjust VLM_CONCURRENCY in .env (default 2, max 4) — higher = faster but more 429 errors. (4) The VLM cache reuses transcriptions from previous runs of the same manga, so re-runs are instant.
Yes! There are 55 voices available across 8 English accents (US, UK, AU, CA, IE, IN, ZA) plus 8 other languages (Japanese, Korean, Spanish, French, German, Portuguese, Hindi, Chinese). Click the speaker icon next to the voice dropdown to preview any voice before starting the pipeline.
By default, videos are stored locally in the data/jobs/{jobId}/output/ directory. If you configure Mega cloud archive (MEGA_EMAIL + MEGA_PASSWORD in .env), finished videos automatically upload to Mega (20 GB free) and the local file is deleted to free disk space. When you watch a video, it's transparently restored from Mega to a 1-hour temp cache and streamed with seek support.
Common causes: (1) The pipeline-service crashed or restarted mid-job — the service auto-requeues stuck jobs on restart, so just refresh the page. (2) Network issues during scraping — try again or pick a different manga. (3) Disk space full — check available space. (4) All VLM providers rate-limited — wait 30-60 min and retry. You can click the Retry button on any failed/stuck job to restart it without re-entering config.
The 'Archive to cloud' button manually uploads a completed video to Mega cloud storage and deletes the local file. This frees disk space while keeping the video accessible — when you click Download or Play, the video is automatically fetched from Mega. Auto-archive is enabled by default (set AUTO_ARCHIVE=false in .env to disable).
Yes! In the Pipeline Configuration, set 'Chapters to process' to the number of chapters you want (or 0 for all). Each chapter is processed sequentially — scraping, transcribing, and rendering one at a time. More chapters = longer processing time (~6 min per chapter). The progress bar shows overall completion across all chapters.
Yes, completely free. The pipeline uses: z-ai VLM (free tier), edge-tts (free, unlimited), ffmpeg (open source), and YOLO panel detection (open source, runs locally). Optional free enhancements: Groq for faster VLM + narration rewriting, Gemini as a second VLM provider, Mega for 20 GB cloud storage. No paid resources are required.
See DEPLOYMENT.md for the full guide. The app can run as a single Docker container (Dockerfile included) on any free Docker host. For Vercel deployment, set PIPELINE_SERVICE_URL to point to your laptop running the pipeline-service (exposed via Cloudflare Tunnel). The Next.js frontend goes on Vercel (free), the database on Turso (free 9 GB), and videos on Mega (free 20 GB) or local storage.
Recent Activity

Ready to create your first recap video?

Search for any manhwa, configure your preferences, and let the AI pipeline handle the rest. It takes about 6 minutes per chapter.