DeOldify is an open-source deep-learning model that colorizes and restores old black-and-white images and film footage. A developer named Jason (GitHub user jantic) started it in 2018. He archived the repository on 19 October 2024, so it receives no more updates.
The model is a generative adversarial network (GAN), in which a generator network makes images and a critic network judges them; here the generator is a pretrained U-Net with self-attention. DeOldify trains it with NoGAN, a method developed in the project that keeps the benefits of GAN training while spending little time on direct GAN training; this removes most glitches and the flicker in video. Three models are offered: artistic and stable for still images, and video for film, which is colorized frame by frame and then rebuilt. The code runs in Google Colab notebooks or on a local machine with the pretrained weights.
DeOldify is free: the code and the pretrained weights are released under the MIT license.