Researchers have introduced the Jacobian Lens Distance (JLD), a novel metric for evaluating perceptual differences between images. Unlike previous methods that rely on human judgments and are sensitive to resolution changes, JLD derives its perceptual geometry from a frozen vision encoder. This approach allows for a fixed metric tensor, fitted quickly on unlabeled images, providing a clear geometric interpretation. JLD demonstrates state-of-the-art performance across multiple perceptual databases, outperforming existing metrics like LPIPS and DISTS, and shows robustness to resolution variations. The method also extends effectively to video quality assessment. AI
IMPACT This new metric could improve image compression, restoration, and generation by providing a more accurate and robust way to evaluate visual quality.
RANK_REASON The item is a research paper detailing a new metric for image perceptual distance. [lever_c_demoted from research: ic=1 ai=1.0]
Read on Hugging Face Daily Papers →
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
