Benran Hu

I'm a Research Engineer at AMD GenAI team, focusing on generative models.

I received my MS in Computer Science at CMU, where I worked with Prof. Ioannis Gkioulekas on 3D reconstruction, inverse rendering, and differentiable rendering. Previously, I interned at Snap Research with Ivan Skorokhodov on building better video generation models. I received my B.Sc. in Computer Science from HKUST, where I was advised by Prof. Chi-Keung Tang, Prof. Yu-Wing Tai, and Prof. Pedro V. Sander.

Email  /  Resume  /  Full CV  /  Scholar  /  GitHub

Research

I'm interested in 3D vision, inverse rendering and 3D reconstruction, scene understanding, generative models, as well as other fields in computer vision and graphics. Currently, I have been working on text-to-image generative models. My past research is mainly on empowering radiance fields and other neural scene representations with the capability of scene understanding. I also work on improving 3D reconstruction and inverse rendering, and accelerating real-time rendering with temporal reprojection.

Text-to-image generation with a 1D binary autoencoder for high compression rates and efficient training/inference.

Aligning spectral properties of RGB and latent spaces helps to create better autoencoders for diffusion models.

Fusing multi-view SAM segmentation masks as an object field improves the performance of zero-shot 3D segmentation in NeRF.

3D instance segmentation in NeRF by matching multi-view instance masks with sparse 3D masks produced by a 3D Mask R-CNN.

We introduce 3D object detection to NeRF by sampling feature grids from radiance fields and applying a 3D detector on them.

This is another website using Jon Barron's template.