UFV-Splatter: Pose-Free Feed-Forward 3D Gaussian Splatting Adapted to Unfavorable Views

Yuki Fujimura, Takahiro Kushida, Kazuya Kitano, Takuya Funatomi, Yasuhiro Mukaigawa
Nara Institute of Science and Technology
arXiv 2025
Paper Code

Abstract

This paper presents a pose-free, feed-forward 3D Gaussian Splatting (3DGS) framework designed to handle unfavorable input views. A common rendering setup for training feed-forward approaches places a 3D object at the world origin and renders it from cameras pointed toward the origin---i.e., from favorable views, limiting the applicability of these models to real-world scenarios involving varying and unknown camera poses. To overcome this limitation, we introduce a novel adaptation framework that enables pretrained pose-free feed-forward 3DGS models to handle unfavorable views. We leverage priors learned from favorable images by feeding recentered images into a pretrained model augmented with low-rank adaptation (LoRA) layers. We further propose a Gaussian adapter module to enhance the geometric consistency of the Gaussians derived from the recentered inputs, along with a Gaussian alignment method to render accurate target views for training. Additionally, we introduce a new training strategy that utilizes an off-the-shelf dataset composed solely of favorable images. Experimental results on both synthetic images from the Google Scanned Objects dataset and real images from the OmniObject3D dataset validate the effectiveness of our method in handling unfavorable input views.


Unfavorable view

Favorable views (red cameras): the object is placed at the world origin, and the cameras are oriented toward the origin. Unfavorable views (green and blue cameras): generated by adding a translation of random direction and magnitude $\rho r$ to each favorable camera, where $r$ denotes the distance from the object origin to the camera. FreeSplatter, trained only on favorable views, performs well when inputs are also favorable but struggles under unfavorable inputs (FreeSplatter w/o recentering).　A naïve approach, such as recentering the foreground to mimic favorable views, is insufficient to resolve these generalization issues (FreeSplatter w/ recentering).　Our proposed method remains robust to unknown and varying camera poses.

unfavorable view

Model architecture

Given a pose-free feed-forward 3DGS model pretrained on favorable views, we aim to adapt it to handle unfavorable views and produce consistent Gaussians across different viewpoints. We adopt FreeSplatter with LoRA layers as our pose-free feed-forward 3DGS backbone. To leverage the priors learned from favorable views, we recenter input images by shifting the foreground regions to the image centers and resizing them accordingly. Additionally, we propose an adapter module that refines the initial Gaussians by predicting residual corrections, using recentered positional embeddings to recover original locations and maintain accurate pixel-to-ray correspondences.

overview

Training

To avoid the costly process of downloading and filtering large-scale 3D assets to generate unfavorable views, we instead use an off-the-shelf, high-quality 2D dataset such as G-Objaverse, which contains only favorable views, for training. We estimate Gaussians using the pretrained model and synthesize training images from unfavorable viewpoints. To compute the loss, we render target images from the estimated Gaussians. However, due to the pose-free input and the recentering process, the predicted Gaussians may differ in scale and field of view across images, and their poses may no longer align with the camera parameters of the target views. To address this, we propose a simple yet effective alignment method for the refined 3D Gaussians, enabling accurate target view rendering during training.

training

BibTeX

@article{Fujimura_2025_UFVSplatter,
    author    = {Fujimura, Yuki and Kushida, Takahiro and Kitano, Kazuya and Funatomi, Takuya and Mukaigawa, Yasuhiro},
    title     = {UFV-Splatter: Pose-Free Feed-Forward 3D Gaussian Splatting Adapted to Unfavorable Views},
    journal   = {arXiv preprint arXiv:2507.22342},
    year      = {2025}
}