Researchers have introduced WildRelight, a new benchmark dataset designed to evaluate single-image relighting models in real-world scenarios. Existing models, often trained on synthetic data, struggle with the complexities of real-world scenes. WildRelight addresses this by providing high-resolution outdoor scenes with aligned, varying natural illuminations and corresponding high-dynamic-range environment maps. The dataset enables a new approach to domain adaptation, allowing synthetic models to adapt to real-world statistics using a physics-guided inference framework that integrates Diffusion Posterior Sampling (DPS) with temporal Sampling-Aware Test-Time Adaptation (TTA). This method transforms the challenging sim-to-real problem into a self-supervised task. AI
IMPACT This dataset and methodology could improve the real-world applicability of AI-powered image relighting techniques.
RANK_REASON The cluster describes a new benchmark dataset and associated research paper for AI-driven image relighting. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
