Title: iLRM: An Iterative Large 3D Reconstruction Model

Original: arXiv:2507.23277v3 Announce Type: replace Abstract: Feed-forward 3D modeling has emerged as a promising approach for rapid and high-quality 3D reconstruction. In particular, directly generating explicit 3D representations, such as 3D Gaussian splatting, has attracted significant attention due to its fast and high-quality rendering. However, many state-of-the-art methods, primarily based on transformer architectures, suffer from severe scalability issues because they rely on full attention across image tokens from multiple input views, resulting in prohibitive computational costs as the number of views or image resolution increases. Toward a scalable and efficient feed-forward 3D reconstruction, we introduce an iterative Large 3D Reconstruction Model (iLRM) that generates 3D Gaussian representations through an iterative refinement mechanism, guided by three core principles: (1) decoupling the scene representation from input images to enable compact 3D representations; (2) decomposing global multi-view interactions into a two-stage attention scheme to reduce computational costs; and (3) injecting high-resolution information at every layer to achieve high-fidelity reconstruction. Experimental results on widely used datasets, such as RE10K and DL3DV, demonstrate that iLRM outperforms existing methods in both reconstruction quality and speed.

Rewrite:

Feed-forward 3D modeling has recently gained traction as a viable solution for achieving swift and superior 3D reconstruction. Specifically, the direct creation of explicit 3D formats, notably 3D Gaussian splatting, has drawn considerable interest owing to its capacity for rapid, high-quality rendering. Nevertheless, leading methods that predominantly utilize transformer architectures face significant scalability challenges. These models depend on full attention mechanisms across image tokens derived from various input viewpoints, which leads to excessive computational burdens when handling increased image resolutions or a greater number of views. To address the need for scalable and efficient feed-forward 3D reconstruction, we present the iterative Large 3D Reconstruction Model (iLRM). This model produces 3D Gaussian representations via an iterative refinement process, underpinned by three fundamental strategies: (1) separating scene representation from input images to facilitate compact 3D structures; (2) breaking down global multi-view interactions into a two-stage attention framework to lower computational expenses; and (3) incorporating high-resolution data at each layer to ensure high-fidelity outcomes. Evaluations on established datasets, including RE10K and DL3DV, confirm that iLRM surpasses current approaches in both the quality of reconstruction and processing speed.

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