albertpumarola/GANimation

Implements a conditional GAN architecture that conditions on Action Unit (AU) annotations extracted via OpenFace, enabling fine-grained control over individual facial muscle activations and their combinations. The approach learns a continuous manifold of anatomically-grounded expressions, allowing magnitude-controlled manipulation of specific AUs rather than discrete emotion categories. Built on PyTorch 0.3.1 with standard training/testing pipelines and dataset preparation utilities for AU annotation preprocessing.

1,985 stars. No commits in the last 6 months.