Origin Lab published a depth study and statistical evaluation on Hugging Face that shows how a model trained only on game-engine data can transfer to real photographs with no fine-tuning and better prediction indoors after fine-tuning.
Published
Signal category
Research & Knowledge
Quote
“A few months ago we did not know whether a model trained only on game-engine data would transfer to real photographs. Now we have an answer with numbers attached.”
— Victor Vielliard|Origin Lab team
Company
Origin Lab
Licensed Multimodal Data for Artificial World Intelligence™
- Industry
- Technology, Information and Internet
- Location
- San Francisco, US
- Company size
- 15 employees
Origin Lab is building the premium data layer for Artificial World Intelligence™: models that understand space, physics, behavior, and how complex environments actually work. We acquire, create, and structure high-fidelity multimodal data from video games, 3D environments, and interactive media, captured at the engine level and enriched with telemetry, depth, action, and scene metadata. All of it fully licensed, with provenance, delivered as queryable datasets built to a lab's specification. We're doing this alongside the AI labs training frontier models, the publishers and developers whose worlds are the most sophisticated simulations ever built, and the academic researchers defining what these systems should understand. Together we turn the world's visual and interactive IP into a licensed foundation for machine intelligence, and the creators behind it participate in the AI economy rather than being written out of it. Backed by a round led by Lightspeed, with SV Angel and Eniac Ventures. Artificial World Intelligence™ is a trademark of Origin Lab.
Founded 2025