Sergey Tulyakov

PhysStream is an autoregressive model for physics-grounded image-to-video synthesis that incorporates structured scene memory and supports fine-grained motion control via sparse velocity-increment signals that encode physical quantities, letting the model learn the underlying dynamics.

EgoPlay achieves a remarkable 17.7% improvement in editing quality over the previous state-of-the-art, redefining the capabilities of egocentric video editing.

Achieving real-time retargeting of human facial motions to non-humanoid avatars, RegHead sets a new standard for animation fidelity and efficiency.

Achieve high-fidelity 4D mesh generation from video by cleverly repurposing existing positional encodings for temporal information, sidestepping the usual trade-offs between expressiveness and pretraining compatibility.

World models can now handle complex multi-agent interactions, controlling up to seven players simultaneously with improved action accuracy and identity consistency.

Achieve up to 40% better FID/FDD scores in diffusion transformers by decoupling image resolution from compute via a simple "elastic latent interface" that lets you dial the latency-quality tradeoff at inference time.

Forget tedious prompt engineering – RefVFX lets you copy and paste visual effects between videos with a single reference clip.