CVPR 2026 · Main Conference

GH-NAF
Grid-Adaptive Hash-Level Attended
Neural Attenuation Fields

Discrepancy-aware sparse-view CBCT reconstruction via frequency-disentangled attention over multi-resolution hash grids.

Seong Je Oh*† Ju Hwan Lee* Chae Yeon Lim* Donghwan Lee Myung Jin Chung Kyungsu Kim†

* Equal contribution  ·  † Corresponding authors

Paper Incoming Code OpenReview Slides Poster

Real CBCT projections carry scatter, beam-hardening, and noise—uniformly fusing multi-resolution hash features blends them into every level. GH-NAF supervises each hash level independently with attenuation-gradient and uncertainty cues, disentangling frequency content so structure survives where naïve fusion fails.

01 Abstract

Existing NeRF-based CBCT methods assume an idealized monoenergetic setting and uniformly fuse multi-resolution hash features, blending projection discrepancies and heterogeneous frequencies into a single representation. We introduce GH-NAF, which trains each hash-grid level independently under uncertainty-weighted supervision. A learned attention selects coarse levels for homogeneous tissue and fine levels at structural boundaries, while a discrepancy-corrected renderer absorbs scatter and beam-hardening at projection time. The result is sharper boundaries, uniform intra-material contrast, and reduced low-frequency bias—on both real and synthetic CBCT.

02 Contributions
i

Hash-Level Attention

Per-point softmax over hash levels replaces uniform concatenation for frequency-aware disentanglement.

ii

Gradient-Guided Selection

Attenuation-gradient soft targets, modulated by uncertainty, teach the attention which scale to trust.

iii

Discrepancy-Aware Rendering

A differentiable corrector absorbs scatter and beam-hardening, keeping them out of the attenuation field.

03 Method

Independent levels, uncertainty-weighted supervision.

GH-NAF pipeline overview
Overview. The hash encoder ℋ with attention gϕ fuses level features into vℓ(x); Fθ predicts (μ, σ, β²) and renders log-projections that match raw measurements. Gradient-derived soft targets Qi(r) align level-wise attention via KL.
04 Results
+0.12 MANIQAchest phantom, 125 v
48.95 dB PSNRFIPS Walnut, 50 v
+2.86 dB over NAFFIPS Walnut, 25 v
13/13 synthetic volstop SSIM @ 50 v
Real-World — Chest Phantom

Pronounced beam-hardening from mobile CBCT. GH-NAF recovers intra-material contrast and parenchymal texture lost by NAF, SAX, R²-Gaussian, Vol3DGS, and even 720-view FDK.

Chest phantom qualitative comparison
Low-Discrepancy Reference — FIPS

Walnut · Pine · Seashell, 25 and 50 views. GH-NAF outperforms NAF, SAX, R²-Gaussian, and Vol3DGS in PSNR/SSIM with sharper boundaries.

FIPS qualitative comparison
Synthetic — Discrepancy-Free

13 volumes simulated with TIGRE. Even without the discrepancy module, hash-level attention alone delivers cleaner tissue and crisper outlines.

Synthetic dataset qualitative comparison
05 Citation
bibtex 
@inproceedings{oh2026ghnaf,
  title     = {GH-NAF: Grid-Adaptive Hash-Level Attended Neural Attenuation
               Fields for Discrepancy-Aware CBCT},
  author    = {Oh, Seong Je and Lee, Ju Hwan and Lim, Chae Yeon and
               Lee, Donghwan and Chung, Myung Jin and Kim, Kyungsu},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision
               and Pattern Recognition (CVPR)},
  year      = {2026}
}