F⁴Splat

Given the same target Gaussian budget, F⁴Splat allocates Gaussians non-uniformly using predicted densification scores, preserving fine details while avoiding redundant allocation in simple or overlapping regions. As a result, it maintains high representation fidelity even with fewer Gaussians and consistently outperforms prior baselines.

Supplementary videos

Video 1

This video provides a more detailed visualization of the adaptive allocation behavior of F⁴Splat. For the samples presented in the main paper, we show the predicted densification score maps and highlight in red the regions where Gaussians are allocated under different Gaussian budgets.

(a). Figure 1 of the main paper.

(b). Figure 3 of the main paper.

(c). Figure 3 of the main paper.

(a). Gaussians are first allocated to spatially complex regions, such as object boundaries and fine edges, indicating that F⁴Splat prioritizes regions that require higher representational fidelity.

(b). Because the upper context image is more spatially complex than the lower one, Gaussian allocation proceeds earlier and more densely in the upper image, indicating that the predicted densification scores reflect image-wise spatial complexity.

(c). In the lower image, the region overlapping with the upper image is assigned low densification scores. This suppresses Gaussian allocation in the shared region and thereby reduces redundant allocation across views.

Video 2

This video compares F⁴Splat with AnySplat on representative scenes from RE10K and ACID. The first row shows the rendered RGB quality, and the second row presents the corresponding depth maps. In the third row, allocated Gaussian locations are highlighted in red. While AnySplat distributes Gaussians uniformly over the scene, F⁴Splat allocates them spatially adaptively. Ultimately, F⁴Splat maintains high fidelity even with fewer Gaussians.

Overview

F⁴Splat is a Gaussian-count controllable feed-forward 3DGS framework for sparse uncalibrated images. Instead of relying on uniform pixel-to-Gaussian or voxel-to-Gaussian allocation, it predicts densification scores and performs densification-score-guided allocation, producing compact yet high-quality 3D representations without retraining for each target budget.

Overview of F⁴Splat. From multi-view context images, the model predicts camera parameters, multi-scale Gaussian parameter maps, and densification score maps, then performs spatially adaptive Gaussian allocation to generate a compact 3D representation under a user-specified Gaussian budget. A single trained model therefore supports different Gaussian budgets at inference time without retraining.

BibTeX

@misc{kim2026f4splat,
  title         = {F4Splat: Feed-Forward Predictive Densification for Feed-Forward 3D Gaussian Splatting},
  author        = {Kim, Injae and Kim, Chaehyeon and Bae, Minseong and Joo, Minseok and Kim, Hyunwoo J.},
  year          = {2026},
  eprint        = {2603.21304},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2603.21304}
}