Colorful Image Colorization

Given a grayscale photograph, this work addresses the problem of hallucinating a plausible color version, which is inherently underconstrained. Rather than relying on user interaction or producing muted results as prior methods did, the authors design a fully automatic system that generates vibrant, realistic colorizations. They embrace the uncertainty of the problem by posing colorization as a classification task over quantized color outputs and use class rebalancing during training to encourage a diverse range of colors. The model runs as a single feed-forward pass through a convolutional neural network at test time and is trained on over a million color images.

Evaluated with a colorization Turing test in which humans choose between generated and real color images, the method fools participants on 32% of trials, significantly outperforming previous approaches. Beyond image colorization itself, the authors show that colorization is a powerful pretext task for self-supervised feature learning, acting as a cross-channel encoder, and it achieves state-of-the-art performance on several feature-learning benchmarks. This dual contribution made it influential for both graphics and representation learning.