Oxford-led Oneira adds a World State Table so video world models remember your edits
jiqizhixin · x · 2026-10-20
Researchers from Oxford, Monash, and Bristol present Oneira (ArXiv 2610.01614), an interactive video world model that fixes a key flaw: existing models act on the wrong objects and forget changes once the camera moves away.
- Root cause: world state lives only in generated frames and context, never recorded separately.
- Approach: a Coding Agent maintains a World State Table recording every object and its state; an engine plans actions, renders a coarse conditioning video along the camera path, and a video generation model produces the segment. Interactions appear as object changes (removal, recoloring); moved objects are re-registered from generated frames once they rest.
- Result: edits persist across the rest of the video, for both input-image objects and newly generated ones.
The accompanying figure shows the full pipeline: agent planning ("pick up the lower left apple"), per-chunk state table updates, and closed-loop segment generation.
More from Multimodal
- Open-source AI workflow turns a 'what if' question into a cinematic film in 10 minutes — w1kke · 2026-10-20
- Free FLUX LoRA turns Google Earth screenshots into realistic drone footage — anthara_ai · 2026-10-20
- Seedance 2.5 wows early testers: AI animation reaches a whole new level — anthara_ai · 2026-10-20
- TischEins Releases Two Free Midjourney Magazines With Prompts and SREF Codes — tisch_eins · 2026-10-20
- AI-generated home video looks near-indistinguishable from real footage — Otherwise_Beyond9481 · 2026-10-20
- Kromoss model updated on Hugging Face, replacing yesterday's version — mikemend · 2026-10-20