SERES (Semantic-Aware Reconstruction from Sparse Views) was jointly proposed in August 2025 by a research team from Shanghai Jiao Tong University, the University of Manchester, and the Chinese University of Hong Kong. The relevant research findings were published in the paper "SERES: Semantic-Aware Neural Reconstruction from Sparse Views".

SERES is a semantically aware sparse view reconstruction framework that enriches the neural field representation with semantic logits, the initial values of which can be obtained from a training-free segmentation model and a visual Transformer. By optimizing these semantic logits, symbolic distance field, and radiation field, reliable feature matching is achieved, resulting in high-fidelity reconstruction. During optimization, geometric primitive masks are also used as regularization, providing additional constraints to mitigate shape ambiguity. SERES successfully reconstructs complex sculptures using only nine viewpoints, preserving accurate geometry and capturing fine details.