jantic/deoldify

DeOldify is an archived deep learning project for automatically adding color to black-and-white photographs and video footage. The repository has been marked as archived by its creator as of October 2024, meaning it is no longer actively maintained. The project's technical approach centers on a training method called NoGAN. Traditional GAN (Generative Adversarial Network) training, where two neural networks compete, one generating images and one evaluating them, produces good colorization but causes flickering artifacts in video. NoGAN runs only a short phase of GAN training (30 to 60 minutes using a small fraction of training data) after first training on the standard task of predicting colors from grayscale. This combination produces stable, consistent colorization without the flickering. Three separate models were developed: an "artistic" model that produces more vivid colors, a "stable" model with fewer glitches, and a video model. Color consistency across video frames is achieved partly through rendering at higher resolutions and through the inherent consistency the models learned during training rather than any explicit temporal modeling. The easiest way to try it is through hosted services linked in the readme. Colab notebooks (cloud-based Jupyter environments) are provided for running colorization in a browser without any local setup. A Desktop application plugin and an in-browser implementation using a different technology are also linked.