flynnmd/deconvfaces

Implements a generative model trained on the Radboud Faces Database with controllable latent space interpolation across identity and emotion attributes. Uses a stacked deconvolution architecture adapted from the 3D object generation literature, scalable to 512x640 resolution through configurable layer depth and kernel counts. Provides multiple generation modes—single, random, smooth interpolation, and keyframe animation—controlled via YAML configuration, built on Keras with NumPy/SciPy for data processing.

892 stars. No commits in the last 6 months.