GAIA-2 — not integrated
GAIA-2 is Wayve’s latent-diffusion world model for controllable multi-camera driving video (successor to GAIA-1). The technical report describes an 8.4B space-time transformer trained with flow matching, a video tokenizer, and structured conditioning (ego dynamics, agents, weather, road semantics). Wayve later discussed a larger GAIA-3 research line.
OpenTryOn status
| Path | Status |
|---|---|
| Hosted API | Not published. Wayve has no public developer API |
| Local weights | Not released |
| Registry | Not added — nothing to call until weights or an endpoint ship |
When a first-party inference API or Hugging Face weights plus a Diffusers/vLLM recipe appear, integrate as Path A (gaia-2) on video-generate and/or Path B (gaia-2-local). Conditioning is driving-specific (multi-camera, ego action), so it is out of the fashion/media default until a usable interface exists.
Do not confuse with unrelated “Gaia” LLM or earth-observation models.