Apple researchers have developed SHARP technology that can generate realistic 3D scene representations from a single photo in under a second on a standard GPU. This technology uses a neural network with a single forward pass to regress 3D Gaussian representation parameters of the scene, supporting real-time rendering of high-resolution, detailed nearby views at over 100fps. Experimental results show that SHARP achieves state-of-the-art effects on multiple datasets, with LPIPS error reduced by 25-34% and DISTS error reduced by 21-43% compared to previous best models, while synthesis time is reduced by three orders of magnitude. The technology has metric properties, supports absolute scale and camera movement, and demonstrates strong zero-shot generalization capabilities, potentially bringing revolutionary applications in AR/VR, autonomous driving, and gaming.
