Apple Unveils LiTo: Reconstructs Complete 3D Objects from a Single Image with View-Consistent Lighting and Shadow Effects

According to CoinWorld, based on monitoring by 1M AI News, Apple’s AI research team published a paper at ICLR 2026 proposing a 3D generation method called LiTo (Surface Light Field Tokenization). It can generate complete 3D objects from a single image and maintain consistent lighting effects such as specular highlights and Fresnel reflections when switching viewpoints. Previously, most 3D reconstruction methods could only handle either geometry or diffuse appearance, making it difficult to restore view-dependent lighting details. LiTo encodes object geometry and view-dependent appearance into the same 3D latent space, then uses a latent flow matching model to generate results from a single image. The training data consists of thousands of 3D objects, each rendered from 150 viewpoints under three lighting conditions. The decoder learns full geometry and appearance reconstruction by randomly sampling sub-sets. Experiments show that LiTo outperforms existing methods like TRELLIS in visual quality and fidelity to input images. The paper was authored by Jen-Hao Rick Chang, Xiaoming Zhao (co-first author), Dorian Chan, and Oncel Tuzel, and is publicly available on arXiv.