Determine whether OX-NeRF gains continue with greater computational resources

Determine whether the reconstruction-quality gains obtained by increasing the number of training scenes and the per-scene hash-table size in OX-NeRF continue on more capable hardware.

Background

OX-NeRF’s ablation study shows that reconstruction quality improves as the number of training scenes and the per-scene hash-table size increase. However, both quantities were capped by GPU-memory limitations, with the hash grids for all scenes kept resident during training and the reported experiments already requiring more than 90 GB of VRAM in the high-view setting. The paper therefore leaves unresolved whether scaling these two resources further would continue to improve reconstruction accuracy on hardware with greater memory capacity.

References

The most immediate follows from the ablation: the number of scenes and the hash table size both raise reconstruction quality and both were capped by memory, so whether the gains continue on more capable hardware is an open question.

— OX-NeRF: 3D X-ray Tomography Reconstruction from Sparse Views Using Implicit Neural Representation  (2610.11547 - Welsch et al., 8 Oct 2026) in Section Conclusions