Point-based representations have recently gained popularity in neural rendering. While they offer many advantages, rendering them from close-up views often results in holes. In splatting-based neural point renderers, these are caused by gaps between different splats, which cause many rays to not intersect with any splat when viewed close-up. A different line of work uses attention to estimate each ray's intersection by interpolating between nearby points. Our work builds on one such method, known as Proximity Attention Point Rendering (PAPR), which learns parsimonious and geometrically accurate point representations. While in principle PAPR can fill holes by learning to interpolate between nearby points appropriately, PAPR also produces holes when rendering close-up, as the intersection point is often predicted incorrectly. We analyze this phenomenon and propose two novel solutions: a method for dynamically selecting nearby points to a ray for interpolation, and a robust attention method that better generalizes to local point configurations around unseen rays. These significantly reduce the prevalence of holes and other artifacts in close-up rendering compared to recent neural point renderers.
(a) Splatting-based renderers: gaps between splats cause holes
(b) Attention-based renderers: PAPR's cylinder selection vs. our dynamic selection (DPS)
In splatting-based methods, gaps between splats cause holes. In PAPR, holes arise because (1) the true intersection may fall outside the convex hull of selected points, and (2) the attention fails to generalize to unseen ray configurations.
Dynamic Point Selection (DPS): Instead of selecting neighborhood points based solely on orthogonal distance to the ray, we use a learned weighted combination of the orthogonal distance and the distance along the ray. The weights are optimized during training to align with the attention mechanism's assessment of point relevance, dynamically adapting the local neighborhood to better capture the true intersection.
Robust Attention: We normalize the ray-dependent point features to maintain consistent scale regardless of camera distance, and introduce a Ray Perturbation Strategy (RPS) that augments ray directions during training. This exposes the attention mechanism to a more diverse range of local point configurations, improving generalization to unseen close-up views.
Ray Perturbation Strategy: the ray endpoint is translated on the image plane while keeping the origin fixed.
| Method | Close-up Synthetic Dataset | Close-up Real Dataset | ||||
|---|---|---|---|---|---|---|
| PSNR ↑ | SSIM ↑ | LPIPS ↓ | PSNR ↑ | SSIM ↑ | LPIPS ↓ | |
| DPBRF | 13.06 | 0.352 | 0.525 | 7.301 | 0.076 | 0.779 |
| SNP | 14.41 | 0.573 | 0.451 | 13.49 | 0.599 | 0.608 |
| Point-NeRF | 19.48 | 0.746 | 0.396 | 13.90 | 0.706 | 0.551 |
| 3DGS | 21.45 | 0.763 | 0.291 | 18.15 | 0.782 | 0.423 |
| Mip-Splatting | 22.96 | 0.786 | 0.286 | 18.97 | 0.794 | 0.404 |
| PAPR | 18.55 | 0.720 | 0.357 | 16.41 | 0.728 | 0.499 |
| Ours | 24.18 | 0.858 | 0.273 | 19.20 | 0.825 | 0.400 |
All methods use 30,000 points. Our method outperforms all baselines on all metrics.
We incrementally add our proposed components to the vanilla PAPR baseline.
Robust attention significantly improves hole-filling. Adding DPS further fills remaining gaps and refines surface geometry.
@inproceedings{zhang2026paprupclose,
title = {PAPR Up-close: Close-up Neural Point
Rendering without Holes},
author = {Zhang, Yanshu and Vashist, Chirag
and Peng, Shichong and Li, Ke},
booktitle = {International Conference on
3D Vision (3DV)},
year = {2026}
}