3rd Place — Gemini 3 Seoul Hackathon (Feb 28, 2026) — 20,000ドル Google Cloud Credits

Turn your wildest fantasy into a full webtoon — starring yourself — in under a minute.

Demo Video

Watch the demo →

In Korean, 망상 (mangstoon) means your wildest, most embarrassing fantasy — the one you replay in your head but would never tell anyone.

MangstoonAI turns that into a scroll-style Korean webtoon. Type your story, upload a selfie, pick an art style, and watch as AI generates a 20+ panel webtoon where you are the main character. Then edit any panel with natural language.

[Story Text + Selfie]
 │
 ▼
[decompose_story] → 22-panel storyboard (5-act structure, JSON)
 │ characters, scenes, dialogue, camera angles, mood
 ▼
[generate_panel ×ばつ 22] → all panels in parallel:
 │ 1. Flash optimizes image prompt
 │ 2. Flash Image generates 9:16 panel
 ▼
[Webtoon Output] → 22 panels in scroll format + speech bubbles
 │
 ▼
[Edit] → "5번 패널 배경을 한강으로" → regenerates just that panel

mangstoon_ai/
├── agent.py # root_agent (Gemini 3.1 Pro, 3 tools)
├── styles.py # 4 art style definitions
├── gcs.py # GCS upload utility
├── tools/
│ ├── story_engine.py # decompose_story() — text → 22-panel storyboard
│ ├── image_gen.py # generate_panel() — 2-step: optimize prompt → gen image
│ ├── panel_editor.py # edit_panel() — 2-step: edit prompt → regen image
│ └── character.py # extract_character() — selfie → character description
└── prompts/
 └── system.py # 3-phase director: Brainstorm → Generate → Edit

• 2-step prompt pipeline: Each panel goes through a Flash prompt optimizer before image generation — narrative paragraphs, not keyword lists
• Face/outfit separation: Face description is permanent identity; outfit varies per scene for visual variety
• Parallel generation: All 22 panels fire simultaneously via asyncio.gather() — under 1 minute total
• Agent + direct call hybrid: Agent handles creative story decomposition, then we break out of the event stream and do image gen in parallel ourselves

• Python 3.12+
• uv (Python package manager)
• Node.js 18+
• A Google AI Studio API key

# Clone
git clone https://github.com/jays0606/mangstoon_ai.git
cd mangstoon_ai
# Backend — set API key
echo "GOOGLE_API_KEY=your_key_here" > backend/.env
# Install Python deps
cd backend && uv sync && cd ..
# Install frontend deps
cd frontend && npm install && cd ..

# Option 1: ADK Dev UI (interactive agent chat)
cd backend && uv run --project .. adk web --port 8080
# → http://localhost:8080/dev-ui/ → select mangstoon_ai
# Option 2: FastAPI backend + Next.js frontend
cd backend && ./run.sh # port 8000
cd frontend && npm run dev # port 3000
# → http://localhost:3000

Returns 22 panels with image_url, dialogue, scene_description, camera_angle, mood, and more.

Returns the regenerated panel with a new image_url.

The app is deployed as:

• Backend: Cloud Run (Seoul / asia-northeast3)
• Frontend: Vercel (auto-deploys on push to main)
• Images: GCS bucket (public read, Seoul region)
• CI/CD: Push to main → GitHub Actions → Cloud Run deploy

To deploy your own instance, set up:

1. A GCP project with Cloud Run, GCS, and Secret Manager
2. Workload Identity Federation for GitHub Actions auth
3. GitHub secrets: WIF_PROVIDER, WIF_SERVICE_ACCOUNT
4. Vercel project linked to the repo with NEXT_PUBLIC_BACKEND_URL env var

Built at the Gemini 3 Seoul Hackathon, Feb 28 2026. Powered by Google ADK + Gemini 3.1 Pro + Gemini 3.1 Flash Image.