Self-Supervised Spontaneous Latent-Based Facial Expression Sequence Generation

In this article, we investigate the spontaneity issue in facial expression sequence generation. Current leading methods field are commonly reliant on manually adjusted conditional variables to direct model generate a specific class of expression. We propose neural network-based method which uses Gaussian noise generation process, removing need for manual control variables. Our takes two sequential images as input, with additive noise, and produces next image sequence. trained types models: single-expression, mixed-expression. With unique movements certain emotion can be generated; mixed expressions, fully spontaneous achieved. compared our current variety publicly available datasets. Initial qualitative results show visually more realistic expressions action unit (AU) trajectories; initial quantitative using quality metrics (SSIM NIQE) generated is higher. approach novel generation, potential wider applications other tasks.