OculoGen: Personalized Text-to-Eye-Movement Video Generation via Cascaded Event-Guided Diffusion

Text-to-video models struggle to follow oculomotor instructions while preserving the anatomy of a reference eye. We present OculoGen, a cascaded framework for personalized text-to-eye-movement video generation from a description and a reference image. OculoParse converts the text into ordered events, OculoKine predicts gaze and eyelid-openness trajectories, and OculoSynth renders them with control-guided diffusion. These explicit intermediate representations allow independent supervision and targeted error diagnosis. A retrieval-and-warp Control Adapter aligns annotated exemplars to the target trajectory, bridging idealized controls and real eye anatomy, while event-conditioned temporal modulation supplies action information to the diffusion model. On held-out TEyeD recordings, OculoGen improves visual fidelity, temporal consistency, and reference-identity similarity over the evaluated baselines and receives higher ratings in a blinded ophthalmologist study. In external evaluations, features extracted from generated videos retain disease-related information learned from real gaze records and improve classification on real held-out data under limited supervision. These results support controllable synthesis as a source of additional training observations for biomedical algorithm development.