Neham Jain

3D generative models · data · evaluation

NehamJain

I build the data and evaluation systems that turn 3D generative models from impressive demos into reliable tools. My work spans data curation, training pipelines, and evaluation — currently at Meshy.ai; previously at Meta Reality Labs, Adobe Research, and CMU.

Get in touch CV Scholar GitHub LinkedIn
SmokeSeer results on a forest scene — left to right: thermal input, smoky capture, desmoked view, and clean 3D reconstruction Thermal input → smoky capture → desmoked → clean reconstruction
3DV 2026 MS Thesis · CMU

SmokeSeer: seeing through smoke with 3D Gaussian splatting

Fuses RGB and thermal video to jointly remove smoke and reconstruct the 3D scene behind it, decomposing wildfire imagery into smoke and surface Gaussians with fluid-dynamics-aware temporal modeling.

Project page Video results Thesis PDF

Overhead multi-camera-stage capture from Embody3D showing four seated people represented with colored 3D body meshes Official Embody3D capture · Meta Reality Labs
Dataset Meta Reality Labs · 2025

Embody3D

A large-scale multimodal dataset for understanding human motion and behavior. I built the QA tooling, automatic quality filters, and curation pipeline behind the release.

1.6Mvideos annotated 500hmotion data curated

Dataset page Paper

About

I’m a Research Scientist at Meshy.ai working on generative models for 3D — multimodal conditioning, training, and the evaluation and data systems they depend on. Before Meshy I was a Research Engineer at Meta Reality Labs, where I shipped the data pipelines behind Embody3D, and I interned twice at Adobe Research, writing CUDA kernels to make relightable Gaussian-splat training 12× faster.

I hold an MS in Robotics from Carnegie Mellon, advised by Prof. Ioannis Gkioulekas in collaboration with Prof. Sebastian Scherer, and a BTech in Electrical Engineering from IIT Madras. I’m always open to research discussions and collaborations — reach me at nehamjain2002 [at] gmail.com.

Highlights