DeOldify

DeOldify is an open-source project that uses deep learning to colorize and restore black-and-white photographs and film frames. It was created by Jason Antic and Dana Kelley and released to make colorization accessible to researchers, educators, and enthusiasts. The project is built in Python with PyTorch and the FastAI library, and it relies on neural networks trained to map grayscale inputs to plausible color images while preserving details and structure.

Its core approach combines convolutional networks and generative adversarial networks (GANs) to produce believable colorization. The repository provides pre-trained models for still images and for video frame colorization, with two main variants: Artistic and Stable. The Artistic variant tends toward more vibrant, painterly results, whereas the Stable variant emphasizes photorealistic color, closer to natural hues.

DeOldify is widely used through the GitHub project, local scripts, and online demos, and has been applied to archival research, media restoration, and education. While it often yields impressive colors, color choices are inferential and may not reflect historical accuracy; artifacts and miscoloring can occur, especially with ambiguous lighting or unusual subject matter. The project remains community-driven, with ongoing updates and tutorials.