World in World: training-free control of frozen video world models for re-camera and revisits
udmrzn · x · 2026-09-13
The arXiv paper 'World in World' introduces a training-free inference-time interface for controlling frozen autoregressive video world models. It converts heterogeneous control evidence — source-video observations, target-view scene projections, geometry renderings, and retrieved generated states — into camera- and time-labelled visual-evidence K/V read through the model's native self-attention. A correspondence router pairs persistent point identities with geometry, while Evidence-wise attention CFG (EWA) regulates each auxiliary channel in a single denoising pass. Capabilities include camera-controlled rerendering, consistent long-horizon revisits, and human-motion tasks without retraining.
More from Multimodal
- Hooking Mage's MCP into Claude generated narrative image sets in one shot — AIandDesign · 2026-10-31
- At 3 am she stops scrolling and a whole town answers in songs — AI short film — ownubie · 2026-10-31
- Phone-scanned homes let anyone tour houses in a browser — and it costs ~$200 per scan — jonstephens85 · 2026-10-31
- AI image demo: the V1 prompt that later made the arrow fly — charis_ai · 2026-10-31
- NVIDIA open-sources NemotronLabs RE-USE, a 9.6M-param multilingual speech denoiser — NVIDIAAI · 2026-10-31
- AI-generated fake trailer imagines a 1999 Dino Crisis movie with retro VHS feel — noizlab_studio · 2026-10-31