Brief

Researchers have introduced HAFMat, a novel framework designed to improve the estimation of physically based rendering (PBR) materials from single human images. This method addresses the inherent ambiguity in such estimations by employing a Multi-layer Adaptive Feature Fusion Mechanism. This mechanism adaptively integrates various guidance cues, including appearance, body geometry, and semantic information, at different stages of the decoding process. Experiments show that HAFMat achieves state-of-the-art results on both synthetic and real-world data for material estimation and subsequent relighting tasks. AI

Researchers have developed PIXLRelight, a novel feed-forward system for controllable single-image relighting. This approach bridges physically based rendering with learned image synthesis by utilizing shared intrinsic conditioning derived from either real photos or PBR renders. The system decomposes images into albedo, diffuse shading, and residuals to condition a transformer-based neural renderer, enabling arbitrary PBR-style lighting control with high image detail preservation. PIXLRelight achieves state-of-the-art quality and operates in under a tenth of a second per image. AI