All credits: https://github.com/jantic/DeOldify

DeOldify is a Black and White image Colorizing library created by Jason Antic. DeOldify introduced the NoGAN technique to solve some of the key problems to make hyper-realistic colorization images and video. DeOldify provides three primary models for different use case.

Demo Video

Image Colorizer.ipynb

Each of these has some limitations and benefit:

This model achieves vibrant color and detailed images, but you have to adjust the parameters a lot to get the best results. You have to adjust the rendering resolution and factors to get the most accurate colorize image.

The model uses a resnet34 backbone on a UNet with an emphasis on the depth of layers on the decoder side. And it is trained on five critic pretrain/GAN cycle repeats via NoGAN.

This model archives the best results in portraits and landscapes. It makes sure nothing got too much colored and makes the most part of the image remain gray like faces and limbs. It is less hyper-realistic but it makes sure nothing looks more colored.

It uses a resnet101 backbone on a UNet with an emphasis on the width of layers on the decoder side.

As the name suggests, it is a model used to color the videos and we are going to see each of these models working in a python environment. It gives smooth, consistent, and flicker-free Video. This model is the same as a stable model in the case of architecture but different in training. DeOldify is trained on 2.2% of Imagenet dataset once at 192px, using only the initial generator/critic pretrain/GAN NoGAN training.