generative-compression

Why we included this project

Anyone reproducing Agustsson et al.'s 2018 GAN paper in TensorFlow will find this repo a close match to the original method. It includes training scripts and a compress.py that loads a trained checkpoint and writes a side-by-side comparison of the original and reconstructed image, which makes the quality loss at extreme bitrates easy to see. This is research code, not a production encoder: it targets TensorFlow 1.8, trains with batch size 1, and the author is upfront that the decoder can hallucinate texture, like greenery showing up inside buildings. That honesty about failure modes, plus the pretrained models and the conditional variant that takes semantic maps, makes it a useful reference for anyone trying to reproduce the paper's results or study the perceptual-quality-versus-bitrate tradeoff.