Sara Ghazanfari

Ph.D. Candidate @ NYU · Incoming Senior Applied ML Researcher @ Apple

I'm a Ph.D. candidate in Electrical and Computer Engineering at New York University, advised by Siddharth Garg and Farshad Khorrami in the EnSuRe Research Group. My research is on how multimodal large language models perceive, align, and reason about visual content, most recently temporal and cross-video reasoning in video MLLMs.

I've interned at Adobe (2025), building evaluation frameworks for visually-guided image editing, and at Apple (2026), working on video temporal grounding. Starting November 16, 2026, I'm joining Apple as a Senior Applied ML Researcher on the Video Apps team.

Google Scholar GitHub LinkedIn X CV

s.ghazanfari.ce [at] gmail [dot] com

News

Oct 2026Released a follow-up to SYNCR on learning cross-video reasoning from simulation.
Sep 2026SYNCR accepted to a NeurIPS 2026 Workshop as an Oral.
Aug 2026Wrapped up my internship at Apple on video temporal grounding.
Jun 2026Presented Chain-of-Frames at CVPR 2026.
May 2026Joined Apple as an Applied Research Scientist Intern.

Show older newsHide older news

May 2026SYNCR released on arXiv, a controlled synthetic benchmark for cross-video reasoning.
Feb 2026Chain-of-Frames accepted to CVPR 2026.
Dec 2025Attended NeurIPS 2025 to present SpotEdit and Chain-of-Frames.
Sep 2025SpotEdit accepted to a NeurIPS 2025 Workshop, with benchmark and code released on GitHub.
Jun 2025Chain-of-Frames released on arXiv, with code and models on GitHub.
May 2025Joined Adobe as a Research Intern, advancing research on unified multimodal models.
May 2025EMMA accepted to TMLR 2025.
Apr 2025UniSim accepted to a CVPR 2025 Workshop.
Jun 2024Attended CVPR 2024.
Jan 2024LipSim accepted to ICLR 2024.
Jun 2023R-LPIPS accepted to an ICML 2023 Workshop.
Jan 2023Started my Ph.D. at NYU.

Publications

SYNCR teaser: paired video frames and segmentation maps

SYNCR: Diagnosing and Learning Cross-Video Reasoning from Simulation

S. Ghazanfari, S. Garg, P. Krishnamurthy, F. Khorrami

Preprint, 2026

A simulator-grounded framework for diagnosing and learning cross-video reasoning. Fine-tuning on SYNCR nearly doubles accuracy and transfers to temporal ordering on real footage.

PDF arXiv

SYNCR: A Cross-Video Reasoning Benchmark with Synthetic Grounding

S. Ghazanfari, S. Garg, P. Krishnamurthy, F. Khorrami

NeurIPS 2026 Workshop · Oral

A controlled synthetic benchmark for cross-video reasoning with programmatically verified grounding: 8,163 multi-video QA pairs built with Habitat, Kubric, and CLEVRER.

PDF arXiv Code

Chain-of-Frames example: reasoning that cites frames 3, 7 and 12

Chain-of-Frames: Advancing Video Understanding in Multimodal LLMs via Frame-Aware Reasoning

S. Ghazanfari, F. Croce, N. Flammarion, P. Krishnamurthy, F. Khorrami, S. Garg

CVPR 2026

Video LLMs whose reasoning steps are grounded in, and explicitly refer to, the relevant frames, outperforming baselines without complex inference pipelines.

PDF arXiv Code

SpotEdit: Evaluating Visually-Guided Image Editing Methods

S. Ghazanfari, W. Lin, H. Tian, E. Yumer

NeurIPS 2025 Workshop

A benchmark for visually-guided image editing that exposes how leading models, including GPT-4o, hallucinate visual cues.

PDF arXiv Code

Towards Unified Benchmark and Models for Multi-Modal Perceptual Metrics

S. Ghazanfari, S. Garg, N. Flammarion, P. Krishnamurthy, F. Khorrami, F. Croce

CVPR 2025 Workshop

UniSim-Bench tracks the progress of perceptual similarity metrics across uni- and multimodal tasks. We propose UniSim, a set of multi-task models that lay the groundwork for general-purpose perceptual metrics.

PDF arXiv Code

EMMA: Efficient Visual Alignment in Multi-Modal LLMs

S. Ghazanfari, A. Araujo, P. Krishnamurthy, S. Garg, F. Khorrami

TMLR 2025

A lightweight module that fuses visual and textual encodings with minimal added parameters, producing instruction-aware visual representations that boost performance and reduce hallucinations.

PDF arXiv Code

LipSim: A Provably Robust Perceptual Similarity Metric

S. Ghazanfari, A. Araujo, P. Krishnamurthy, F. Khorrami, S. Garg

ICLR 2024

We show that state-of-the-art perceptual similarity metrics are vulnerable to adversarial attacks and propose LipSim, trained with 1-Lipschitz networks, a robust metric with provable guarantees.

PDF arXiv Code

R-LPIPS: An Adversarially Robust Perceptual Similarity Metric

S. Ghazanfari, S. Garg, P. Krishnamurthy, F. Khorrami, A. Araujo

ICML 2023 Workshop

We show LPIPS is sensitive to adversarial perturbation and propose R-LPIPS, which uses adversarially trained deep features to be robust where the standard metric is not.

PDF arXiv Code