UniqueSplat: View-conditioned 3D Gaussian Splatting for Generalizable 3D Reconstruction

Traditional feed-forward 3D Gaussian Splatting (3DGS) methods typically predict a fixed set of Gaussians for a scene, which are then used to render all novel views. This approach often fails to account for the specific geometric and appearance requirements of individual viewpoints, leading to artifacts like blurring or cracks. UniqueSplat addresses this by asking: can we design a feed-forward model that dynamically adapts its Gaussian representation to the specific target view query?

The authors propose a two-branch view-conditioned hypernetwork that integrates view-specific information into the Gaussian prediction process. The architecture consists of two primary components:

1. View-Agnostic Branch: This branch maintains learnable embeddings that capture shareable, scene-wide knowledge, ensuring consistency across different viewpoints.
2. View-Specific Branch: This branch projects features from input sparse-view images into the target viewpoint's space. It then generates an embedding that adapts the model's parameters to the specific query view.

By fusing the weights generated by these two branches, the model produces customized Gaussian parameters—including center, opacity, covariance, and color—for each individual view, allowing for more precise and realistic rendering than static models.

UniqueSplat demonstrates superior performance in novel view synthesis across several benchmark datasets, including RealEstate10K, ACID, and DTU. The dual-branch design effectively disentangles general scene knowledge from view-specific requirements, resulting in higher-quality depth maps and reduced visual artifacts compared to state-of-the-art methods like MVSplat. Furthermore, the model exhibits strong cross-dataset generalization, confirming its ability to adapt to unseen scenes without requiring per-scene optimization.