F4Splat: Feed-Forward Predictive Densification for Feed-Forward 3D Gaussian Splatting
Injae Kim, Chaehyeon Kim, Minseong Bae, Minseok Joo, Hyunwoo J. Kim
2026-03-24
Summary
This paper introduces a new technique called F4Splat for creating 3D models from images, focusing on making the process faster and the resulting models more efficient.
What's the problem?
Existing methods for quickly building 3D models from images often create too many unnecessary details, especially in simple areas, and struggle to control the overall complexity of the model. They essentially spread out 3D 'blobs' (Gaussians) without considering where they're actually needed, leading to wasted resources and potentially blurry results. It's hard to tell the system 'I want a model with roughly this many details' without having to start the whole process over.
What's the solution?
F4Splat solves this by intelligently deciding where to place these 3D blobs. It predicts a 'densification score' for each part of the scene, indicating how much detail is required. Areas with lots of complexity get more blobs, while simple areas get fewer. This allows for a more focused and efficient use of resources, and importantly, lets users specify a desired number of blobs for the final model without needing to retrain everything.
Why it matters?
This research is important because it allows for the creation of high-quality 3D models much faster and with less data than previous methods. By reducing the number of unnecessary details, it makes the models more manageable and improves the quality of images generated from new viewpoints. This has implications for applications like virtual reality, augmented reality, and creating 3D content more easily.
Abstract
Feed-forward 3D Gaussian Splatting methods enable single-pass reconstruction and real-time rendering. However, they typically adopt rigid pixel-to-Gaussian or voxel-to-Gaussian pipelines that uniformly allocate Gaussians, leading to redundant Gaussians across views. Moreover, they lack an effective mechanism to control the total number of Gaussians while maintaining reconstruction fidelity. To address these limitations, we present F4Splat, which performs Feed-Forward predictive densification for Feed-Forward 3D Gaussian Splatting, introducing a densification-score-guided allocation strategy that adaptively distributes Gaussians according to spatial complexity and multi-view overlap. Our model predicts per-region densification scores to estimate the required Gaussian density and allows explicit control over the final Gaussian budget without retraining. This spatially adaptive allocation reduces redundancy in simple regions and minimizes duplicate Gaussians across overlapping views, producing compact yet high-quality 3D representations. Extensive experiments demonstrate that our model achieves superior novel-view synthesis performance compared to prior uncalibrated feed-forward methods, while using significantly fewer Gaussians.