Avinash Paliwal

I am a Founding Research Scientist at Nuance Labs, where we build visual conversational AI that feels human.

I received my PhD from Texas A&M University, where I was advised by Nima Kalantari in the Aggie Graphics Group. My research focused on 3D Computer Vision and Computational Photography, with an emphasis on 3D scene reconstruction and generation.

Previously, I was a Founding AI Researcher at Morphic, specializing in 3D Computer Vision and Generative Video, focusing on controllable generation using techniques like 3D Gaussian Splatting, large diffusion models, and 3D/4D reconstruction. Before that, I was a Research Scientist Intern at Meta Reality Labs (2023), a Research Intern at Leia Inc. (2021), and a SDE Intern at Amazon (2019).

Email  /  CV  /  Google Scholar  /  GitHub  /  LinkedIn

Two independent random-walk crops of one monocular video become a source-target pair, turning internet-scale footage into free supervision for camera-controlled reshooting of dynamic scenes.

Posing single-image 360° generation as joint panorama-and-depth optimization, solved by alternating minimization, keeps the lifted 3D scene globally coherent.

Two specialized diffusion priors, one repairing visible regions and one inpainting unseen ones, reconstruct a detailed 3D scene from just a handful of images.

Tying sparse-view Gaussians to an implicit decoder keeps them coherent, avoiding the usual "jumble of needles" when only three images are available.

Matching content and style Gaussian distributions with a Wasserstein-2 (Earth-Mover's) distance transfers scene detail directly, with no generative latent-space losses.

Reshading the image before relocating its pixels lets specular highlights travel with the camera instead of staying glued to surfaces.

Decomposing wide stereo baselines into multi-plane disparities, while warping time non-uniformly, enables joint view-and-time interpolation of stereo video.

Encoding bidirectional flow into a coordinate network via a hypernetwork yields continuous intermediate flows that stay robust even when brightness changes between frames.

Aligning and fusing frames in stages avoids distant-frame misalignment, while a gradient-mask-conditioned discriminator stops the GAN from hallucinating noise in smooth regions.

A low-res, high-speed auxiliary camera supplies the true motion to reconstruct 1080p slow-motion video, capturing non-linear motion that interpolation alone misses.

The original website!