Transforming real or generated videos to 4D scenes with a multi-view diffusion model.

Using video models to simulate scene variations and training a multiview diffusion model to harmonize inconsistent captures.

Leveraging a vision-language model to infer procedural materials from input images.

A video-to-video model that synthesizes large-angle novel viewpoints of dynamic scenes.

Enabling interaction with static 3D objects by distilling material parameters from video generation models.

Building a multi-view conditioned diffusion model and using it as a prior to regularize radiance field reconstruction from only a few images.

A diffusion model trained on a single 3D textured shape, enabling generating diverse high-quality variations of it.

Finetuning the stable diffusion model on synthetic 3D data to enable zero-shot novel view synthesis from a single image of an object.

Solving time-dependent PDEs by evolving an implicit neural spatial representation over time.

Training a multi-scale patch GAN on a single example to generate 3D shapes that locally resemble the input.

A realtime speech denoising system with a dynamic sliding window.

Using transformers to encode and decode parametric CAD modeling sequences.

Denoising speech signals by recovering noise from "silent" time periods.

Using conditional GANs to complete a partial 3D scan with multiple plausible outputs.

Using a sequence-to-sequence network to generate 3D shapes via sequential part assembly.

Transferring the video-captured motion from one performer to another by learning a character-agnostic motion representation.