S2R: Illumination-Aware Synthetic-to-Real Adaptation for Single-Image Human Material Estimation

Single-image human material estimation aims to recover physically based rendering (PBR) attributes of a human from a single image, enabling realistic relighting for immersive multimedia applications. Although recent methods have achieved promising results on synthetic benchmarks, their predictions on real images may suffer from residual illumination artifacts, which lead to unrealistic relighting results. We argue that this limitation mainly arises from the domain gap between synthetic and real data in both material appearance and illumination, which aggravates the ambiguity between illumination and reflectance during inference.

To address this issue, we propose S2R, a novel two-stage synthetic-to-real framework for single-image human material estimation. In the first stage, we train the model on high-quality synthetic data with supervision under multiple constraints, and introduce an illumination-aware feature injection module to improve the disentanglement of material properties from shading. In the second stage, we selectively fine-tune the model on real images to transfer the knowledge learned from synthetic supervision to the real domain. This adaptation process is guided by self-supervised reconstruction and teacher-student regularization, which together reduce the synthetic-to-real domain gap while preserving stable material estimation.

Extensive experiments on both synthetic and real-image benchmarks demonstrate that our method produces cleaner material estimates with less residual illumination and achieves more realistic relighting results than existing approaches.