DiffSSMEye: Conditional High-Frequency Eye Movement Synthesis with Structured State-Space Diffusion
Synthesizing high-frequency eye movements requires preserving long-range temporal dependencies and subject-specific oculomotor dynamics. However, methods based on generative adversarial networks (GANs) and variational autoencoders (VAEs) can struggle to preserve these characteristics simultaneously. We introduce DiffSSMEye, a conditional denoising diffusion probabilistic model with bidirectional Structured State Space Sequence (S4) layers for generating five-second gaze-velocity sequences at 1,000 Hz. Generation is conditioned on a coarse gaze reference, user identity, and task, where identity and task embeddings are incorporated through feature-wise linear modulation. Identity-preservation, multi-resolution spectral, and integrated-drift losses complement noise prediction to constrain biometric consistency, frequency structure, and displacement error. Subject-disjoint evaluation on GazeBase across seven tasks yields mean real--synthetic embedding cosine similarities of 0.91--0.95. DiffSSMEye also achieves the lowest median spatial errors among these baselines for horizontal and random saccade tasks. These results demonstrate improved spatial accuracy and biometric feature consistency for subject-specific gaze synthesis.
Authors
- Mehedi Hasan Raju (ORCID: https://orcid.org/0000-0002-1144-6118)
- Oleg V. Komogortsev (ORCID: https://orcid.org/0000-0001-7890-8842)
- Kamrul Hasan (ORCID: https://orcid.org/0000-0002-1844-3569)
Publication Details
- Journal
- arXiv (Cornell University)
- Published
- 2026-11-12
- DOI
- https://doi.org/10.48550/arxiv.2611.17013
- Primary Topic
- Gaze Tracking and Assistive Technology
- Type
- preprint