ImageBind
ImageBind One Embedding Space to Bind Them All
ImageBind is a multimodal embedding framework that learns a shared representation space across six modalities—images, text, audio, depth, thermal, and IMU (inertial motion) data—without requiring explicit pairwise training for every modality combination. Instead of aligning each pair independently, ImageBind uses image data as the central binding modality, aligning all other modalities to it so they can interoperate zero-shot. This creates a unified embedding space where representations from...