LVM introduces UVDv1, a large-scale dataset built specifically for training purely visual large models. Its key innovation is the "visual sentence," a unified format designed to solve the challenge of heterogeneous visual data from diverse sources. Specifically, the dataset integrates five major categories of data totaling 1.64 billion images/frames — including unannotated images, annotated images, videos, annotated videos, and 3D synthetic objects. All of this data, whether single images, video frames, or image-annotation pairs, is converted into a uniform sequence format: a visual sentence consisting of one or more images followed by an end-of-sentence token. This allows the model to perform autoregressive next-token prediction over visual sequences in the same way language models process text, enabling a large vision model to be trained on massive, diverse, purely visual data — without any language labels — and to perform multiple tasks via prompting.