Answers are grounded in this paper and cite their sources.
Gamut mapping is a crucial component of the image signal processing (ISP) pipeline and color conversion between devices. Improper gamut mapping can introduce noticeable hue shifts and artifacts, resulting in poor visual quality. Although recent studies have significantly improved the fidelity of gamut extension through learning-based models and efficient metadata encoding schemes, the quality of gamut reduction and the intrinsic relationship between gamut reduction and extension have been largely overlooked. In this paper, we focus on developing deep learning-based approaches for gamut mapping that produce images with high color fidelity and perceptual quality for bidirectional conversion between wide gamut and small gamut. We propose an invertible framework that performs both gamut reduction and gamut extension within a single model. In the forward pass, the input wide-gamut image is decomposed into a clipped component and a residual component, which are jointly transformed into a small-gamut image and a latent variable. In the reverse pass, the small-gamut image and a Gaussian latent variable are used to reconstruct the wide-gamut image. To ensure that gamut reduction preserves hue and perceptual color similarity to the wide-gamut image, we introduce a self-supervised hue loss. We conduct experiments on three different gamut mapping scenarios: ProPhoto RGB to sRGB, BT.2020 to sRGB, and sRGB to a smaller "Toy" gamut. The proposed method outperforms previous metadata-free approaches in both color fidelity and perceptual quality across the evaluated gamut mapping scenarios.
No takes yet. Share an insight, caveat, or question.
Xue et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: