Distilling Style from Image Pairs for Global Forward and Inverse Tone Mapping

TL;DR

This paper introduces a method to extract and encode style information from image pairs into a low-dimensional, interpretable space, improving the accuracy of global tone mapping tasks over existing approaches.

Contribution

It proposes a novel style distillation technique using normalizing flows conditioned on polynomial color basis, outperforming PCA and VAE in encoding style for tone mapping.

Findings

Achieves nearly 40 dB accuracy in style encoding.

Outperforms state-of-the-art methods by 7-10 dB.

Provides an interpretable 2-3D style representation.

Abstract

Many image enhancement or editing operations, such as forward and inverse tone mapping or color grading, do not have a unique solution, but instead a range of solutions, each representing a different style. Despite this, existing learning-based methods attempt to learn a unique mapping, disregarding this style. In this work, we show that information about the style can be distilled from collections of image pairs and encoded into a 2- or 3-dimensional vector. This gives us not only an efficient representation but also an interpretable latent space for editing the image style. We represent the global color mapping between a pair of images as a custom normalizing flow, conditioned on a polynomial basis of the pixel color. We show that such a network is more effective than PCA or VAE at encoding image style in low-dimensional space and lets us obtain an accuracy close to 40 dB, which is about 7-10 dB improvement over the state-of-the-art methods.