RealRep: Generalized SDR-to-HDR Conversion via Attribute-Disentangled Representation Learning

TL;DR

This work tackles SDR-to-HDR conversion under real-world degradations by learning attribute-disentangled representations that separately model luminance and chrominance. It introduces RealRep, a framework consisting of multi-view degradation encoders, a fusion module, and a degradation-domain aware controlled mapping network (DDACMNet) that uses degradation priors to guide adaptive HDR reconstruction. The approach employs contrastive learning to enforce degradation-invariant, attribute-separated embeddings and a two-stage training strategy to stabilize learning. Experiments on HDRTV4K and HDRTV1K demonstrate superior generalization and perceptual fidelity, highlighting RealRep's potential for robust HDR color gamut reconstruction in practical applications.

Abstract

High-Dynamic-Range Wide-Color-Gamut (HDR-WCG) technology is becoming increasingly widespread, driving a growing need for converting Standard Dynamic Range (SDR) content to HDR. Existing methods primarily rely on fixed tone mapping operators, which struggle to handle the diverse appearances and degradations commonly present in real-world SDR content. To address this limitation, we propose a generalized SDR-to-HDR framework that enhances robustness by learning attribute-disentangled representations. Central to our approach is Realistic Attribute-Disentangled Representation Learning (RealRep), which explicitly disentangles luminance and chrominance components to capture intrinsic content variations across different SDR distributions. Furthermore, we design a Luma-/Chroma-aware negative exemplar generation strategy that constructs degradation-sensitive contrastive pairs, effectively modeling tone discrepancies across SDR styles. Building on these attribute-level priors, we introduce the Degradation-Domain Aware Controlled Mapping Network (DDACMNet), a lightweight, two-stage framework that performs adaptive hierarchical mapping guided by a control-aware normalization mechanism. DDACMNet dynamically modulates the mapping process via degradation-conditioned features, enabling robust adaptation across diverse degradation domains. Extensive experiments demonstrate that RealRep consistently outperforms state-of-the-art methods in both generalization and perceptually faithful HDR color gamut reconstruction.

Figures (6)

• Figure 1: An illustration of our motivation. (a) Comparison between previous frameworks (top) and our attribute-disentangled method (bottom), which explicitly separates luminance and chrominance and injects them into the learned prior space. (b) Distribution shifts of luminance and chrominance across degradations, motivating the need for disentangled modeling to ensure robustness. (c) t-SNE visualization of SDR input features (left), features from previous methods (middle), and our method (right), all trained on a multi-degradation dataset. Our approach achieves superior attribute separation, enabling better generalization across diverse SDR conditions.
• Figure 2: The superiority of our approach in recovering luminance and chrominance under unknown degradations. (a) Input SDR image. (b) Results of LSNet Guo2023, trained on a single degradation type, produces results that closely mimic the input SDR. (c) Results of ICTCPNet Huang2023VideoIT with multi-degradation training, which still fail to recover realistic brightness and color. (d) Our proposed method (RealRep) achieves significantly better recovery of both luminance and chrominance, demonstrating strong generalization to unseen degradations. (e) Ground truth HDR image.
• Figure 3: Overview of the proposed multi-degradation SDR-to-HDR framework. Our architecture consists of a pair of multi-view degradation encoders that separately extract global and local representations of luminance and chrominance, a multi-view fusion module that facilitates spatial and channel-wise interaction, and the Degradation-Domain Aware Controlled Mapping Network (DDACMNet), which performs adaptive mapping and detail refinement conditioned on degradation-aware features. The encoders are jointly trained with a disentanglement objective to ensure cross-domain consistency. Our framework enables dynamic luminance expansion and perceptually faithful color gamut reconstruction across a wide range of real-world SDR degradations.
• Figure 4: Illustration of (a) Dense Controlled Mapping (DCM) and (b) Sparse Controlled Mapping (SCM).
• Figure 5: Visual comparison with the state-of-the-art methods. Following hdrtvdm Guo2023, the HDR images are shown in their original encoding without tone mapping to preserve highlights. Our method delivers visually superior results, with clearer highlight recovery, more vivid colors, and better structure preservation compared to prior methods. Visualization is best viewed on an HDR screen.