Apple’s LiTo AI Creates Detailed 3D Models with Lighting Effects from One Photo

Researchers from Apple’s machine learning division have developed a new neural network called LiTo, capable of reconstructing three-dimensional objects from a single photograph. The model’s distinctive feature is its ability to convey not just shape, but also complex optical properties: highlights, reflections, and changes in a material’s appearance depending on the viewing angle.

According to Apple, existing neural networks for 3D generation are primarily focused on geometric accuracy. Even when they are capable of generating textures, they typically work with simplified (“flat”) colors and fail to account for how an object interacts with light from different angles.

LiTo tackles this problem differently: the model encodes two parameters simultaneously in its latent space—geometry and optical variability. The network is first trained on the fundamental task of building 3D shapes. Then, by compressing random samples of the surface’s light field into a compact set of vectors, it learns to accurately reproduce visual effects.

In tests, the developers compared LiTo with TRELLIS, another contemporary model. According to Apple, their solution demonstrates higher accuracy and detail during generation. Looking ahead, the technology is intended for use in creating digital assets and rapidly transferring real-world objects into digital 3D models.