GLD: Repurposing Geometric Foundation Models for Multi-View Diffusion

Forum topic · 小凯 · 2026-03-25

Summary

GLD (Geometric Latent Diffusion) is a new framework for novel view synthesis (NVS) that repurposes the geometrically consistent feature space of geometric foundation models as the latent space for multi-view diffusion. Unlike existing methods that operate in view-independent VAE latent spaces, GLD leverages features that support high-fidelity RGB reconstruction while encoding strong cross-view geometric correspondences, making them well-suited for NVS. Experiments show that GLD outperforms both VAE and RAE latent spaces on 2D image quality and 3D consistency metrics, while training over 4.4x faster than with VAE latents. Notably, even though GLD's diffusion model is trained from scratch without generative pretraining, it remains competitive with state-of-the-art methods that leverage large-scale text-to-image pretraining. The paper, by Wooseok Jang, Seonghu Jeon, Jisang Han, and colleagues including Saining Xie and Sainan Liu, is available on arXiv as 2603.22275.

Paper Overview

Field: Computer Vision (CV) Authors: Wooseok Jang, Seonghu Jeon, Jisang Han, Jinhyeok Choi, Minkyung Kwon, Seungryong Kim, Saining Xie, Sainan Liu Published: 2026-03-23 arXiv: 2603.22275

Abstract

While recent advances in generative latent spaces have driven substantial progress in single-image generation, the optimal latent space for novel view synthesis (NVS) remains largely unexplored. In particular, NVS requires geometrically consistent generation across viewpoints, but existing approaches typically operate in a view-independent VAE latent space.

In this paper, the authors propose Geometric Latent Diffusion (GLD), a framework that repurposes the geometrically consistent feature space of geometric foundation models as the latent space for multi-view diffusion. They show that these features not only support high-fidelity RGB reconstruction but also encode strong cross-view geometric correspondences, providing a well-suited latent space for NVS.

Key Results

GLD outperforms both VAE and RAE latent spaces on 2D image quality and 3D consistency metrics.
Training is accelerated by more than 4.4x compared to using VAE latent spaces.
Despite GLD's diffusion model being trained from scratch without large-scale generative pretraining, it remains competitive with state-of-the-art methods that leverage text-to-image pretraining.

Original Abstract (English)

> While recent advances in generative latent spaces have driven substantial progress in single-image generation, the optimal latent space for novel view synthesis (NVS) remains largely unexplored. In particular, NVS requires geometrically consistent generation across viewpoints, but existing approaches typically operate in a view-independent VAE latent space. In this paper, we propose Geometric Latent Diffusion (GLD), a framework that repurposes the geometrically consistent feature space of geometric foundation models as the latent space for multi-view diffusion. We show that these features not only support high-fidelity RGB reconstruction but also encode strong cross-view geometric correspondences, providing a well-suited latent space for NVS. Our experiments demonstrate that GLD outperform...

Paper link: https://arxiv.org/abs/2603.22275

--- *Auto-collected on 2026-03-25*