TP-GAN

Why we included this project

One face photo at an awkward angle is all TP-GAN needs to produce a frontal view of the same person. It comes from the ICCV 2017 paper that introduced the idea, combining a global generator with separate local patches for the eyes and mouth so the result stays photorealistic and the identity holds up. That makes it a natural fit for face recognition and re-identification pipelines that want pose normalization before feature extraction, and the architecture itself is still worth studying. Be aware it is an early TensorFlow 0.12 research release: setup expects a Matlab cropping script and keypoints from off-the-shelf landmark detectors, and use is limited to noncommercial academic work under the CMU license. For frontalization research it works well as a baseline and reference implementation, just not as a turnkey production service.