Eva Lorenzo Iglesias

Generating synthetic time series data, such as videos, presents a formidable challenge as complexity increases when it is necessary to maintain a specific distribution of shown stages. One such case is embryonic development, where prediction and categorization are crucial for anticipating future outcomes. To address this challenge, we propose a Siamese architecture based on diffusion models to generate predictive long-duration embryonic development videos and an evaluation method to select the most realistic video in a non-supervised manner. We validated this model using standard metrics, such as Fréchet inception distance (FID), Fréchet video distance (FVD), structural similarity (SSIM), peak signal-to-noise ratio (PSNR), and mean squared error (MSE). The proposed model generates videos of up to 197 frames with a size of 128×128, considering real input images. Regarding the quality of the videos, all results showed improvements over the default model (FID = 129.18, FVD = 802.46, SSIM = 0.39, PSNR = 28.63, and MSE = 97.46). On the coherence of the stages, a global stage mean squared error of 9.00 was achieved versus the results of 13.31 and 59.3 for the default methods. The proposed technique produces more accurate videos and successfully removes cases that display sudden movements or changes.