Building foundation models for the physical world.

An OR video diffusion framework that synthesizes routine and rare operating-room events from abstract geometric representations to support ambient surgical intelligence.

Diffusion-based world models fine-tuned on laparoscopic footage to simulate the biomechanics of robotic surgical suturing with high temporal fidelity.

Distills an offline dual-stream audio-visual diffusion model into a real-time streaming generator running at ~25 FPS with synchronized multimodal output.

A training-free framework for lip synchronization and audio-visual editing that stabilizes the editing trajectory by removing stochastic elements.

A watermarking framework that cryptographically binds audio and video latents in joint generation models, blocking deepfake swap attacks with >99% integrity.

A zero-shot, training-free image-driven video editing method for rectified flow models that modulates injection intensity using an editing-aware mask.

A propagation-based video editing pipeline that learns from on-the-fly supervision generated by pre-trained video diffusion models — no paired data needed.

An efficient video inpainting control framework that focuses compute on masked tokens, making localized edits 10x cheaper without sacrificing quality.

A few-shot, sim-to-real video diffusion pipeline that animates realistic, expressive hair motion from a single human image using lightweight LoRA modules.

A unified Qwen-VL + LTX framework for next-scene prediction, trained with a causal consistency reward to anticipate plausible futures.

Learns appearance-independent motion representations from raw video, enabling open-world motion transfer between semantically unrelated entities.

A real-time DiT talking-head framework using a temporal VAE and Speech Autoencoder for efficient, well-aligned audio-driven synthesis.

A fast DiT-based video inpainting method using a Circular Position-Shift strategy for strong long-term temporal consistency on large masks.

A training-free omnimatte approach using pre-trained video diffusion models to decompose videos into object layers and effects at real-time speed.

An all-in-one video creation and editing framework that unifies reference-to-video generation, video-to-video editing, and masked editing through a single Video Condition Unit interface.

Activates the in-context generation ability of large video models to produce multiple viewpoint-consistent videos of the same shared world.

The first open platform that benchmarks generative world models inside closed-loop environments, measuring whether they actually help embodied agents succeed.

A world model that pairs panoramic video generation with an evolving explicit 3D memory for spatially consistent long-horizon exploration.

Decouples motion from appearance to adapt generalist video diffusion models into controllable driving world models with under 6% of prior compute.

Unifies video generation and motion planning by feeding the video world model's latents directly into a diffusion planner for consistent driving trajectories.

Uses video diffusion models to synthesize realistic 3D camera and scene-motion variations, augmenting scarce datasets such as UAV imagery.