Vufinder's Map-Anything AI Handles 12+ 3D Reconstruction Tasks in One Model

Vufinder has released Map-Anything, an AI model built on research from Meta and Carnegie Mellon University that performs over 12 types of 3D reconstruction tasks in a single feed-forward pass. The model uses a transformer architecture to recover metric 3D scene geometry and camera parameters directly from one or more images, without requiring prior camera calibration. It supports a wide range of applications including monocular depth estimation, structure-from-motion, depth completion, and camera localization through a unified interface. For large datasets, a memory-efficient inference mode allows the model to scale to hundreds or thousands of views while reducing GPU memory usage. However, accuracy may vary in specialized domains not well-represented in training data, and high-resolution details can be lost as inputs are resized to a maximum of 518 pixels.