Research map / 3D vision / 3D generation
XScale-NVS: Cross-Scale Novel View Synthesis with Hash Featurized Manifold
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes hash featurized manifold, a hash-encoded surface featurization rasterized from an MVS mesh with a deferred MLP neural shader, plus the GigaNVS cross-scale dataset for large-scale novel view synthesis.
Problem
High-fidelity cross-scale novel view synthesis of real-world large-scale scenes, where explicit surface representations lose fine detail (discretization resolution, UV distortion) and volumetric neural fields suffer from dispersed weights and surface ambiguity.
Approach
Featurize the MVS mesh surface with volumetric multi-resolution hash encoding, querying hash features only at rasterized surface intersections so hash entries concentrate on multi-view consistent surface regions; add surface multisampling and latent manifold deformation, decoded by a light MLP neural shader.
Key results
- On GigaNVS, mean LPIPS 0.148 vs 0.243 for ZipNeRF (about 40% reduction) and mean PSNR 20.84 vs 20.09.
- On Tanks&Temples, PSNR 29.66 and LPIPS 0.079 versus 28.67 and 0.110 for iNGP.
- Ablation: removing manifold deformation or multisampling lowers PSNR from 23.02 to 22.77 and 22.24 respectively on TW-Pavilion (Day).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use hash featurized manifold with deferred neural rendering | XScale-NVS · GigaNVS | LPIPS: 0.243 -> 0.148 | Helped |
| use hash featurized manifold with deferred neural rendering | XScale-NVS · Tanks&Temples | LPIPS: 0.114 -> 0.079 | Helped |
| add manifold deformation (ablation) | XScale-NVS · TW-Pavilion (Day) | PSNR: 22.77 -> 23.02 | Helped |
| add manifold deformation (ablation) | XScale-NVS · TW-Pavilion (Day) | SSIM: 0.775 -> 0.786 | Helped |
| add manifold deformation (ablation) | XScale-NVS · TW-Pavilion (Day) | LPIPS: 0.131 -> 0.113 | Helped |
| add surface multisampling (ablation) | XScale-NVS · TW-Pavilion (Day) | PSNR: 22.24 -> 23.02 | Helped |
| add surface multisampling (ablation) | XScale-NVS · TW-Pavilion (Day) | SSIM: 0.752 -> 0.786 | Helped |
| add surface multisampling (ablation) | XScale-NVS · TW-Pavilion (Day) | LPIPS: 0.140 -> 0.113 | Helped |
| reduce mesh resolution to 1M primitives (ablation) | XScale-NVS · TW-Pavilion (Day) | PSNR: 23.02 -> 22.84 | Hurt |
| reduce mesh resolution to 1M primitives (ablation) | XScale-NVS · TW-Pavilion (Day) | SSIM: 0.786 -> 0.782 | Hurt |
| reduce mesh resolution to 1M primitives (ablation) | XScale-NVS · TW-Pavilion (Day) | LPIPS: 0.113 -> 0.114 | Hurt |
Surprising findings
- A 1M-face mesh (90% decimated) yields nearly the same quality as the 10M-face mesh (PSNR 22.84 vs 23.02), suggesting weak dependence on mesh resolution.
Limitations
- Cannot handle incompleteness and occlusions caused by incorrect geometry.
- Relies on an off-the-shelf MVS mesh as the surface proxy.
Experimental setup
Datasets: GigaNVS Tanks&Temples
Benchmarks: GigaNVS Tanks&Temples
Metrics: PSNR SSIM LPIPS training time rendering time GPU memory
Baselines: Meta Representation 3D Gaussian Splatting ZipNeRF Neuralangelo iNGP
Builds on: multi-resolution hash encoding deferred neural rendering Meta Representation ZipNeRF
Closest papers by meaning
- MegaSynth: Scaling Up 3D Scene Reconstruction with Synthesized Data (CVPR 2025 · similarity 0.85)
- Novel View Synthesis with Pixel-Space Diffusion Models (CVPR 2025 · similarity 0.85)
- UMAMI: Unifying Masked Autoregressive Models and Deterministic Rendering for View Synthesis (NeurIPS 2025 · similarity 0.85)
- NAMVIS: Next-Scale Autoregressive Multi-View Image Synthesis (NeurIPS 2026 · similarity 0.85)
- LRM-Zero: Training Large Reconstruction Models with Synthesized Data (NeurIPS 2024 · similarity 0.85)
- LVSM: A Large View Synthesis Model with Minimal 3D Inductive Bias (ICLR 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.