# Open Fern

> Open source, model-agnostic AI video studio. Create short clips, cinematic long-form videos, podcasts, audiobooks, and ads from a single idea.

Open Fern is an open-source (MIT) AI production platform that runs as a self-hosted Docker Compose stack. It is model-agnostic: users can swap AI providers for every stage of the pipeline without changing application code.

## What it does

- **Short-form video** — 15–60 s animated clips generated from a text idea, with optional brand voice and music
- **Long-form video** — 5–30 min image-slideshow productions with AI-written scene scripts, character consistency, and subtitle burn
- **Podcast** — multi-speaker audio episodes with natural AI voices (Inworld AI, ElevenLabs, Chatterbox)
- **Audiobook** — upload a PDF/TXT/EPUB; AI assigns distinct voices to every character and narrates the full text
- **Ads & Branding** — asset-first brand workspace: build reusable products, recurring characters, and shot assets, then launch campaigns
- **Music video** — upload an MP3; AI transcribes lyrics and generates a synced visual video
- **Thumbnails & Titles** — generate 4 click-worthy thumbnail variants and title options for any production
- **Trend Feed** — ingest trending topics from Google Trends, Reddit, Twitter/X, and TikTok; launch a production directly from a trend card

## Technical stack

- Backend: Python 3.12, FastAPI, SQLAlchemy (SQLite default, PostgreSQL-ready), Alembic
- Queue: ARQ (async task queue over Redis)
- Storage: MinIO (S3-compatible, self-hosted)
- Frontend: React 18, Vite, TypeScript, TanStack Query, Zustand
- Deployment: Docker Compose (single VPS, ~$20/month)

## AI providers used

| Stage | Providers |
|---|---|
| Script / LLM | Claude (Anthropic), GPT-4o (OpenAI), Gemini (Google) |
| Image generation | Gemini Flash Image, GPT-Image 1 |
| Video generation | Veo 3 (via Kie.ai), Kling 3.0 (via Kie.ai) |
| Voice synthesis | ElevenLabs, Inworld AI, Chatterbox, OpenAI TTS |
| Music | Suno AI |

## Key design decisions

- All six pipeline modules are pure async functions with typed I/O — no database access inside modules
- SSE (Server-Sent Events) for real-time job progress (unidirectional, proxy-friendly)
- UUID string primary keys for seamless SQLite → PostgreSQL migration
- Asset storage uses canonical key schema: `brands/{id}/...`, `jobs/{id}/...`
- Stage outputs are checkpointed to object storage after each successful call — partial productions resume from the last successful stage

## License

MIT. Source available on GitHub.

## Pages

- [Landing page](https://openfern.ai/)
- [Technical Whitepaper](https://openfern.ai/whitepaper)