Full Professor, Massachusetts Institute of Technology
6 papers at NeurIPS 2025
One-Shot Adaptive Visual Tokenizer
Feed-forward dynamic 3DGS scene reconstruction from videos.
We improve the quality of generative models by using low-quality, corrupted, and out-of-distribution data
Automated agent with a self-reflection mechanism that detects unintended visual attribute dependencies in vision models
We distill large datasets to just one image per class and use them to train linear probes on top of self-supervised vision models.
