Westlake AGI Lab's 'World in World' Steers Frozen Video Models

The framework adds no new training. Westlake AGI Lab's World in World paper bolts fresh camera control, rendered geometry, and long-term memory onto the publicly released causal-fast checkpoint of LingBot-World 2.0 with 'all pretrained parameters frozen,' and its guidance step 'introduces no additional network function evaluations (NFE) for guidance.'

The setup is spare. The backbone's native self-attention already reads what the paper calls 'clean visual states': the initial observation and recently finalized outputs, indexed by camera pose and event time. World in World converts everything else — source-video observations, target-view projections, rendered geometry, retrieved history — into that same representation and lets the model pick it up through attention. The native block retains 'the backbone's 18 latent-frame-state slots: six source anchors, eight recent-history states, and four states in the current chunk.' Temporary auxiliary attention blocks carry extra evidence for the current chunk and are removed once it is finalized.

Two mechanisms decide what the model actually reads. Correspondence-guided attention routing tracks points 'with persistent identities in the source video' and uses depth and camera information to steer queries toward geometrically matched source-video tokens. Evidence-wise attention CFG then strengthens 'only the complementary direction introduced by the evidence relative to the native response,' rather than amplifying the native output along with it.

The strongest number in the paper is in the ablation. Removing target-view warping 'causes the largest degradation across all reported metrics, with rotation and translation errors rising to approximately 3.4× and 10.9× their full-method values.' The full method is reported as achieving the lowest camera trajectory errors and the best or tied-best result on all seven reported VBench dimensions on DAVIS, evaluated against ReCamMaster, TrajectoryCrafter, WorldForge, InSpatio-World, UniWorld-View, and CameraAnything. Camera trajectories are estimated with Depth Anything 3 and ViPE, then averaged across the two.

Beyond rerendering, the same interface is used for bullet-time generation, video stabilization, video editing, human motion transfer, and 'K/V sharing between two generation cases produced by the same frozen model.' The paper does not publish per-baseline VBench scores or wall-clock cost in the passages available. It lands in a busy stretch for video world-model preprints; a neighbouring paper the day before argued the opposite tack, decoupling scene state from the video renderer rather than routing everything through frozen attention.