Staff AI Research Scientist, Meta Superintelligence Labs
I build unified multimodal foundation models that bridge understanding, generation, and editing across text, image, video, and audio. I have trained these models at every stage: pretraining, mid-training, and post-training.
At Meta, I work on post-training for Muse Image and Muse Video, and I led mid-training and post-training of the unified generation and editing model behind the Add Me and Remix Me features in Meta AI. Before Meta, I worked on pretraining for Adobe Firefly Image 5 and Amazon Titan Image Generator (Nova Canvas), and on document understanding and facial analysis at AWS AI.
I received my Ph.D. from the University of Maryland, College Park, advised by Prof. Rama Chellappa, and interned at Waymo, Adobe Research, Palo Alto Research Center, and Siemens Healthineers.
Meta's image and video generation models. I work across the end-to-end post-training stack: supervised fine-tuning, distillation, reward modeling, and reinforcement learning.
Arena rank at launch
Where my work sits in each model's training pipeline.
Staff AI Research Scientist
Post-training for Muse Image and Muse Video. Led mid-training and post-training of a unified model for text-to-image generation, image editing, style transfer, face swap, and person insertion, which powers Add Me and Remix Me in Meta AI.
Senior Applied Scientist
Pretraining for Firefly Image 5, including sparse Mixture-of-Experts and sparse-attention architectures. Led a unified generation and editing framework for image and video, and researched multimodal LLMs for unified understanding and generation.
Senior Applied Scientist; previously Applied Scientist
Pretraining for Titan Image Generator, launched on Amazon Bedrock at AWS re:Invent 2023. Built the core models for Amazon Textract key-value pair detection and document forgery detection, and contributed to Amazon Rekognition face liveness, facial attributes, and face recognition.
Electrical and Computer Engineering, advised by Prof. Rama Chellappa
Research internships at Waymo, Adobe Research, Palo Alto Research Center, and Siemens Healthineers.
One model for a broad range of image generation and editing tasks with diverse input and output configurations, producing realistic shadows, reflections, lighting effects, and pose changes.
Recasts referring image segmentation as polygon generation: a sequence-to-sequence transformer reads image patches and text tokens and predicts polygon vertices autoregressively.
Pair-wise self-consistency learning (PCL) trains ConvNets to extract source features that expose deepfake images.
A landmark-guided attention branch discards features from occluded regions, and a facial-region branch predicts the expression from each facial block independently.
Photo-realistic facial expression editing, with an expression controller that adjusts intensity continuously from low to high.
Localizes attribute-specific face regions with weak supervision and classifies attributes without face alignment.
A two-stage training algorithm that uses a face recognition network to regularize expression recognition trained on small datasets.
Reviewer for CVPR, ICCV, ECCV, NeurIPS, ICLR, ICML, IJCV, and IEEE T-PAMI.