The Many Senses of Visual Similarity: A Text-Prompted Image Perceptual Metric
The episode discusses a paper introducing "The Many Senses of Visual Similarity: A Text-Prompted Image Perceptual Metric." The paper addresses how similarity has many aspects, moving beyond single metrics like LPIPS. The authors built a dataset and metric that allows users to query similarity based on specific visual senses like color or pose, showing significant gaps in current models and offering new ways to evaluate generative models.