World-model data studio for AV long-tail scenarios: compose a scene, let a
world model dream it, drive it (or let autopilot), and download training
samples (video + frame-aligned actions.jsonl in a .tar).
Live: https://d-arnav.github.io/DreamTale-AV/
npm install
npm run dev # :5173Keys go in the API Keys page (in-memory, per tab) or private/.env for dev:
GROQ_KEY, GEMINI_API_KEY, REACTOR_KEY.
| stage | engine | cost |
|---|---|---|
| prompts | Groq gpt-oss-120b |
free |
| seed frame | gemini-2.5-flash-image |
~$0.04/image |
| world model | reactor/lingbot-world-2 (WebRTC, live) |
33 cr/s |
The world stage bills per connected second; every exit path stops the session
(npm run check asserts it). Deeper notes: docs/structure.md.
hover over the logo