GenVOG-DiT: A Transformer-Based Diffusion Model for Pose-Driven, Patient-Agnostic Nystagmus VOG Video Generation

Aimon Rahman, Kemar E. Green, Vishal M. Patel

Keywords: Video Diffusion, Medical AI, Ophthalmology, Neurology

Abstract: GenVOG-Dit is a diffusion transformer-based framework designed to generate realistic synthetic videos of nystagmus, a condition marked by involuntary, repetitive eye movements that compromise visual acuity and often signal underlying neurological disorders. Progress in deep learning for nystagmus analysis has been limited by the scarcity of publicly available data, as eye movement patterns can be personally identifiable, raising privacy concerns. GenVOG-Dit addresses this challenge by generating high-fidelity synthetic nystagmus videos that emulate diverse clinical waveform types without relying on real patient data. Leveraging publicly accessible datasets and a transformer-enhanced diffusion process, the model produces extended, clinically meaningful video sequences. The utility of these synthetic videos is validated through performance on downstream tasks using real patient datasets, demonstrating the potential of GenVOG-Dit as a privacy-preserving solution for nystagmus research.

Primary Subject Area: Image Synthesis

Secondary Subject Area: Generative Models

Registration Requirement: Yes

Visa & Travel: Yes

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 183