Demystifying Real-World Image Colorization with AI

Paper Title:

BigColor: Colorization using a Generative Color Prior for Natural Images

Published on:

20 July 2022

Primary Category:

Computer Vision and Pattern Recognition

Paper Authors:

Geonung Kim,

Kyoungkook Kang,

Seongtae Kim,

Hwayoon Lee,

Sehoon Kim,

Jonghyun Kim,

Seung-Hwan Baek,

Sunghyun Cho

Bullets

Key Details

•

Proposes BigColor, a new automatic colorization method using a learned generative color prior

•

Uses an encoder-generator network to focus the model on color synthesis given spatial structure

•

Enlarges representation space compared to GAN inversion approaches for complex real images

•

Achieves state-of-the-art performance and robustness, especially for in-the-wild images

•

Supports multi-modal colorization and arbitrary input resolutions

Explore the topics in this paper

colorization

conditional generation

generative models

high resolution

multimodal learning

AI generated summary

Demystifying Real-World Image Colorization with AI

This paper proposes BigColor, a novel method to automatically colorize grayscale images, even complex real-world ones, with vivid and natural colors. It uses a deep neural network with an encoder-generator architecture to learn a generative color prior focused on color synthesis. This allows it to handle diverse in-the-wild images beyond the limited representation space of previous GAN inversion approaches. Experiments demonstrate BigColor significantly outperforms state-of-the-art methods.