SHARP is a neural network model for monocular view synthesis that generates 3D Gaussian scene representations from single images.
The tool addresses the challenge of creating photorealistic novel views from a single photograph. It works by regressing parameters of a 3D Gaussian representation through a single feedforward neural network pass, completing the process in under a second on a standard GPU. The resulting 3D Gaussian representation can then be rendered in real time at high resolution for nearby viewpoints. The approach produces metric-scale output that supports absolute camera movements and demonstrates strong zero-shot generalization across different datasets.
Developers should consider SHARP for applications requiring fast, single-image view synthesis with photorealistic quality. The tool is particularly suited for projects where inference speed and generalization across diverse inputs matter more than per-scene optimization. The 3D Gaussian splat output format is compatible with various public renderers, making integration into existing pipelines straightforward. Video rendering with camera trajectories is available but currently requires a CUDA GPU, whereas Gaussian prediction itself runs on CPU, CUDA, and MPS devices.
The project maintains an active research implementation with clear documentation and automated model downloading. Development includes a command-line interface for both prediction and trajectory rendering, with straightforward installation via standard Python package management. The codebase acknowledges multiple open-source dependencies and provides separate licensing terms for code and released models.
