Visual Layer

Visual Layer provides a platform for managing and curating large-scale visual datasets used in training AI models. The company's technology is built on a CPU-only graph engine that processes images, videos, objects, and semantic concepts to uncover connections within datasets ranging from gigabytes to petabytes. The platform embeds data into a multimodal vector space and constructs a graph called the VL Index to expose relationships across datasets without requiring external metadata. Visual Layer's core technology is based on fastdup, an open-source package created by the company that helps identify and fix issues in visual datasets. The platform addresses the problem of incorrectly labeled, broken, missing, or duplicate images that can degrade AI model quality. The company found that up to 30% of image and video collections, amounting to hundreds of millions of assets, fall into this problematic category. For example, their research revealed that the ImageNet-21K pre-training dataset contained over a million pairs of duplicates among its 14 million images.