Collecting eye-movement data from autistic children is challenging, yet such data is needed to develop models for early detection of autism, motivating synthetic data generation to mitigate scarcity constraints. This work proposes a slightly enhanced variant of the SP-EyeGAN model which we call SynEM-GAN. We compare both models that are trained on the GazeBase video data then fine-tuned on a small autism dataset. Synthetic sequence quality is assessed using Jensen-Shannon divergence between generated and real distributions for fixation and saccade features (velocity, acceleration, amplitude, dispersion), aggregated over 10 folds. Results show complementary strengths: SynEM-GAN achieves lower divergence on fixation velocity and mean velocity, whereas the SP-EyeGAN achieves lower divergence on four of five saccade metrics. Across metrics, both models outperform statistical and VAE baselines. The fine-tuned variants underperformed their non-fine-tuned counterparts in this setup. However, they show better results than the baselines on almost every metric, even though the baselines were not adjusted for the autism dataset. These findings support the feasibility of adversarial generators for synthetic eye-movement data in settings with limited autism samples.