Tingting Liao

A causal, streaming version of Code World Model: each 0.71 s chunk is generated from the start image, the proxy so far and the previous chunks, with no access to the future. Diffusion forcing with a noisy history slows long-horizon drift, and a 4-step distilled model renders a chunk in 5.1 s on one H200.

Trains streaming world models on minutes-long spans at bounded cost: only the last chunks are scored, and each retrieves a small bank of co-visible frames instead of attending over the whole history, so a place looks the same when revisited.

A hybrid memory for camera-controlled world models: half the blocks keep a key-value cache of past views, the other half carry a compact recurrent state addressed by camera geometry, so revisited places come back faithfully at bounded memory.

Massive activations in video DiTs concentrate on first-frame and latent-boundary tokens. Steering them at inference improves temporal coherence — training-free, with no extra compute.

A benchmark that asks generative models to reason rather than render: abstract puzzles, embodied navigation and physical commonsense, scored across five reasoning abilities.

Places characters from different sources into a single video, keeping each one's identity and style intact while they interact.

Turns one portrait — photo, painting or cartoon — into a rigged, animatable 3D head, whatever the style.

Generates animatable 3D avatars from text alone, with high-quality geometry and texture.

Reconstructs a fully clothed human from a single image, using text guidance to hallucinate the regions the camera never saw.

Recovers a detailed clothed avatar from one image, combining an implicit surface with parametric body priors.