Title: LagerNVS: Latent Geometry for Fully Neural Real-time Novel View Synthesis

Abstract: While recent advancements demonstrate that neural networks can execute 3D tasks like Novel View Synthesis (NVS) without relying on explicit 3D reconstruction, we contend that incorporating strong 3D inductive biases remains beneficial for network architecture. To substantiate this, we present LagerNVS, an encoder-decoder framework for NVS that leverages "3D-aware" latent features. The system’s encoder is initialized using a 3D reconstruction network that has undergone pre-training with explicit 3D supervision. This component is combined with a lightweight decoder and optimized end-to-end through photometric loss functions. LagerNVS delivers state-of-the-art results for deterministic feed-forward Novel View Synthesis—achieving a Peak Signal-to-Noise Ratio (PSNR) of 31.4 on the Re10k dataset—regardless of whether camera parameters are known. The model operates in real time, demonstrates robust generalization to unconstrained, in-the-wild data, and supports generative extrapolation when integrated with a diffusion-based decoder.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC