apple/ml-sharp: Sharp Monocular View Synthesis in Less Than a Second

SHARP is Apple's open-source implementation for photorealistic view synthesis from a single image. It generates 3D Gaussian representations in under a second via a single neural network pass, enabling real-time rendering of high-resolution images from nearby viewpoints. The system achieves state-of-the-art results with 25-34% LPIPS reduction and 21-43% DISTS reduction compared to previous models, while being three orders of magnitude faster. The release includes a CLI tool for prediction and rendering, with automatic model checkpoint downloading and support for various 3D Gaussian splatting renderers.