Research

Projects

My research combines optics and deep learning to build cameras that protect privacy and authenticity at capture time, and more recently LLM agents and generative models for editing and verifying visual content. Every project below is backed by a paper; see Publications for the full list.

NOWA pipeline: a phase mask in the camera embeds a null-space watermark used to localize tampering
CVPR 2026 · OralRice University

NOWA: Null-space Optical Watermark

A phase mask in the camera aperture hides a watermark in the null space of the optics: invisible in the photo, but it fingerprints the capture and localizes tampering. Validated on a prototype camera.

Paper →arXiv →
Traditional camera versus privacy-preserving camera with an optimized phase mask
CVPR 2024First author

Privacy-Preserving Optics for Face De-identification

Camera optics learned jointly with a network so that identity is removed at capture time, before the image is digitized, while the image stays useful for downstream vision. Core of my PhD.

  • CVPR 2024 (first author)
  • Learning to Describe Scenes via Privacy-Aware Designed Optical Lens, IEEE TCI 2024
  • Optics lens design for privacy-preserving scene captioning, ICIP 2022 Best Paper
Project page →Paper →arXiv →
SGA overview: symbolic geometric verification of LLM-generated animation code
EMNLP 2026First author

SGA: Geometric Verification for Educational Video

A plug-and-play agent that checks LLM-generated Manim code for spatial overlaps on a symbolic scene graph and fixes them before rendering. Raters preferred it over the raw output in 84% of comparisons.

Project page →arXiv →
MambaStyle inversion and editing results on faces and cars
arXiv 2025First author

MambaStyle: Efficient StyleGAN Inversion

A single-stage encoder built on vision state-space models for real-image inversion and editing, with fewer parameters and faster inference than state-of-the-art encoders.

arXiv →
Learned color-coded apertures mounted on camera lenses
IEEE TCI 2024ECCV 2024 Workshops

Coded Apertures for Depth Estimation

Color and binary coded apertures learned end-to-end with a depth network, recovering depth from a single coded image or a low-cost stereo pair, tested on lab prototypes.

TCI paper →ECCVW paper →
People

Collaborations

  • UISHDSP Group, Universidad Industrial de Santander (Colombia). Primary affiliation, Prof. Henry Arguello.
  • KAUSTKAUST (Saudi Arabia). Prof. Bernard Ghanem and Dr. Carlos Hinojosa: SGA, MambaStyle, CVPR 2024.
  • RiceRice University (USA). Computational Imaging Lab, Prof. Ashok Veeraraghavan: NOWA.