Title: CanonCGT: Reference-Based Color Grading via Canonical Pivot Representation

Abstract

The objective of reference-based color grading is to replicate the lighting and tonal atmosphere of a target image while maintaining the original scene’s structural integrity and color harmony. However, current approaches—ranging from filter-based techniques to photorealistic models—frequently suffer from unstable tone mapping. These methods often result in unnatural outputs due to excessive color shifting or inconsistent retention of hues. To address these limitations, we introduce CanonCGT, a novel two-stage framework anchored by a "canonical pivot," which serves as a style-neutral intermediate representation to ensure stable color mapping. In the first stage, the input image is canonicalized by eliminating its inherent tonal bias. The subsequent stage then applies color grading to align the image with the reference style. Our training methodology, DP-CGT, employs a dual-phase scheme that integrates supervised learning from presets with self-supervised refinement using unpaired photographs. Extensive evaluations across various datasets demonstrate that CanonCGT achieves superior visual fidelity and tonal consistency compared to existing state-of-the-art methods. The source code is publicly accessible at https://github.com/Jinwon-Ko/CanonCGT.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC