Disentangles space and time in video diffusion for controllable generative rendering. Given a single video, freely steer camera viewpoint and temporal motion across the 4D space-time domain.
Project Page →
Led the cross-org effort to ship free-form 3D camera control in the Firefly Video Model. Adaptor-based conditioning adds only 3% runtime overhead. T2V shipped Aug 2025; I2V shipped Dec 2025.
Try it live →
Achieving 3D consistency in video generation through geometric priors, enabling cameras to move freely through generated scenes.
Project Page →
A large-scale 4D dataset of motion-captured human-object interactions, enabling research in human motion understanding and generation.
Project Page →
Repurposes pretrained image diffusion models for zero-shot video editing, enabling consistent frame-to-frame style and content manipulation.
Project Page →
Regresses 3D human pose and shape from an RGB image using metric and semantic attributes. Best paper finalist (33 / 8,000 submissions).
Project Page →
A dataset with videos of people in natural scenarios paired with ground-truth 3D body pose/shape and 3D scene scans for dense human-scene contact labeling.
Project Page →