Oneira: interactive video world model uses a CodingAgent-maintained WorldStateTable
机器之心 · wechat · 2026-10-13
Researchers from Oxford, Monash and Bristol propose Oneira, an interactive video world model that separates "what happens in the world" from "how it's rendered." A CodingAgent maintains a WorldStateTable recording each object's label, 3D BBox and interaction state; an Engine renders it into a conditional video that a VideoGenerationModel turns into realistic frames. New objects discovered as the camera moves are automatically registered, and state changes persist over long videos. On their InteractionBench (100 interaction + 50 memory tests), Oneira beats the strongest baselines by 29/19/29 points on interaction occurrence, timing and target accuracy, and improves PSNR/SSIM by 93%/83% on memory tests. Paper: arXiv 2610.01614; code open-sourced.
More from Multimodal
- Flying car concept v2: Grok Imagine stills + MiniMax H3 video generation — creatoroff · 2026-10-16
- Flying car short made with Grok Imagine, Minimax H3 and Topaz upscale — creatoroff · 2026-10-16
- Fish Audio's open-weight Drama 3 voice model claims top emotive quality across 83 languages — mhdfaran · 2026-10-16
- Claude Opus 5.5 makes a pure-code Halloween animation starring bots trick-or-treating for tokens — sattvi · 2026-10-16
- 1,290-line commit lands MOSS Image support in ostris/ai-toolkit — ostrisai · 2026-10-16
- Ostris AI Toolkit adds training support for OpenMOSS's MOSS-Image-SFT model — ostrisai · 2026-10-16