Image embeddings

Image embeddings are compact, fixed-length numerical vectors (e.g., 512-dimensional) that encode an image's high-level semantic content, enabling efficient comparison and machine understanding.

This is the core technology: translating high-dimensional image data into a compact, fixed-length vector—a numeric 'fingerprint' that captures the image's semantic meaning, not just raw pixels. Deep learning models (like CNNs, Vision Transformers, or multimodal systems like OpenAI's CLIP) generate these vectors. Crucially, visual similarity translates directly to vector proximity in the embedding space: close vectors mean similar images. This vector representation is the backbone for high-performance computer vision tasks: powering rapid nearest-neighbor search, clustering, e-commerce visual search, and zero-shot classification on multi-million-image datasets.

https://voxel51.com/docs/fiftyone/glossary/image_embeddings.html