90210: Running the Show Without Property Tax

A new open-source project called 90210 turns a screenplay into a finished short film — complete with synchronized video, native audio, dialogue, music, and subtitles. It runs locally, uses a Next.js PWA frontend and FastAPI backend, and orchestrates Google Veo 3.1, Nano Banana, Gemini 2.5 Pro, and ElevenLabs Music.

How It Works

The pipeline starts with a screenplay. Gemini 2.5 Pro acts as the "Director", interpreting the script and generating a shot list. Veo 3.1 renders video and audio for each shot. Nano Banana handles portrait generation. ElevenLabs Music produces background tracks. All outputs are assembled into a final film.

Quality Control with Oracles

Every quality decision is mathematically grounded. A composite Quality Oracle uses VideoScore, ViCLIP, LAION aesthetic, MUSIQ, ArcFace, LPIPS, and Whisper to score each shot. Low scores trigger auto re-rolls and adaptive tier escalation (e.g., switching to a higher-quality model). A Story Oracle ensures narrative coherence using Reagan-6 arcs, TRIPOD turning points, Average Shot Length, and SBERT coherence.

Quick Start

# one-time setup
python3 -m venv .venv
.venv/bin/pip install -e ./backend
cp .env.example .env  # fill in GEMINI_API_KEY and ELEVENLABS_API_KEY

# start backend
./run-backend.sh  # uvicorn at http://localhost:8000

# frontend (new shell)
cd frontend
pnpm install
pnpm dev  # → http://localhost:3000

Optional: install heavy ML dependencies for the Quality Oracle (~3GB):

.venv/bin/pip install -e "./backend[oracles]"

Architecture

The backend is a FastAPI server that orchestrates API calls to Google and ElevenLabs. The frontend is a Next.js PWA that can be installed to the iOS home screen for a native-like experience. The project includes a design spec, a Veo 3.1 deep-dive, and a cost/latency/quality math script.

Costs

A 2-minute movie costs roughly $20: $0.10–$0.40 per second of video from Veo 3.1, plus $0.01 per second of music from ElevenLabs. The project explicitly documents these costs in docs/superpowers/research/cost_latency_quality_math.py.

Why This Matters for Developers

This is a production-grade, open-source pipeline that combines state-of-the-art generative models with automated quality control. The use of multiple metrics (VideoScore, ViCLIP, etc.) for auto re-rolls is a practical pattern for any AI-powered content generation system. The project also demonstrates how to glue together disparate APIs (Google, ElevenLabs) into a coherent workflow.

Technical Details

Next Steps

Clone the repo, fill in your API keys, and run it. The documentation includes a full design spec and research notes. If you're building AI pipelines, the oracle pattern and cost math are worth studying.