GANimation

Why we included this project

GANimation is the official implementation of an ECCV 2018 paper that animates a single face photo into controllable expressions without any video of the person. Instead of driving motion with a target face or landmarks, it conditions a generative adversarial network on Action Unit annotations, which describe anatomical facial movements on a continuous scale. That lets you set how strongly each unit activates and combine several at once, so one still image can be pushed through different smiles and surprise while the subject's identity stays consistent. The repo is research code rather than a polished product. You prepare your own image set with OpenFace action unit annotations, train, then run a test script to generate new expressions. Researchers reproducing the method and engineers prototyping face-editing pipelines will find it useful, though the pinned PyTorch 0.3.1 stack and hand-built dataset format assume someone comfortable adapting research code rather than expecting plug-and-play.