Research map / Probabilistic and scientific modeling / Variational and generative sampling
Accelerating Neural Field Training via Soft Mining
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.
Accelerates neural field training by soft hard-mining of sampling locations, weighting per-sample loss by a power of the error-based importance and sampling locations with Langevin Monte Carlo.
Problem
Neural field training is slow, and the common uniform or heuristic selection of training pixels/rays yields sub-optimal convergence.
Approach
Reformulate neural field batch construction as importance sampling, then relax hard mining into a soft version by weighting each sample's loss by stop-gradient(importance)^(-alpha) with alpha in [0,1]; sample locations efficiently via Langevin Monte Carlo driven by the error gradient.
Key results
- Reaches 35 dB on 2D image fitting almost 4x faster than uniform sampling (about 2,400 vs 10,600 iterations).
- On LLFF, reaches 25 dB in about 1,700 iterations versus about 3,800 for uniform sampling and 3,300 for Edge.
- Less than 4% added per-iteration time (257 s vs 248 s for 20k iterations on RTX 3090).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use soft mining instead of uniform sampling | NeRF with NerfAcc · LLFF test views, 1k iterations | PSNR: 21.00 -> 24.04 | Helped |
| Use soft mining instead of uniform sampling | NeRF with NerfAcc · NeRF Synthetic test views, 1k iterations | PSNR: 26.50 -> 28.57 | Helped |
| Use soft mining instead of uniform sampling | NeRF with NerfAcc · LLFF test views, 10k iterations | PSNR: 25.92 -> 26.53 | Helped |
| Use soft mining instead of uniform sampling | NeRF with NerfAcc · NeRF Synthetic test views, 10k iterations | PSNR: 30.27 -> 31.38 | Helped |
| Use edge-based sampling instead of uniform sampling | NeRF with NerfAcc · LLFF test views | PSNR: 21.00 -> 21.13; 25.27 -> 25.35; 25.92 -> 25.97 | Mixed |
| Use soft mining instead of uniform sampling | Instant-NGP image fitting · Pluto, Eiffel Tower, and Tokyo images | PSNR: ≈2,400 iterations -> ≈10,600 iterations | Helped |
| Add uniform samples to the soft-mining batch (ablation) | Soft mining on LLFF · LLFF, 1k iterations | PSNR: 23.91 -> 24.04 | Helped |
| Add sample re-initialization (ablation) | Soft mining on LLFF · LLFF, 1k iterations | PSNR: 23.57 -> 24.04 | Helped |
| Set the soft-mining parameter to α=0 (hard mining) (ablation) | Soft mining · LLFF | PSNR | Hurt |
| Use an intermediate soft-mining α instead of pure hard mining or pure importance sampling (ablation) | Soft mining · NeRF Synthetic and LLFF | PSNR | Helped |
| Replace LMC with multinomial sampling from Q(x) (ablation) | Soft mining on the LLFF Room scene · Room | PSNR | Helped |
Surprising findings
- Straightforward hard mining (alpha=0) does not improve, and actually degrades, LLFF convergence.
- The Edge heuristic gives large gains on NeRF Synthetic but not on LLFF real-world scenes.
Limitations
- Relies purely on the training loss, which may not correlate with final rendering quality for NeRF.
- Multinomial sampling ideal is computationally impractical; LMC only approximates it.
- Hyperparameters (alpha, LMC step sizes) are tuned per dataset type.
Experimental setup
Datasets: Eiffel Tower image Pluto image Tokyo image NeRF Synthetic LLFF
Benchmarks: NeRF Synthetic LLFF 2D image fitting (Eiffel Tower, Pluto, Tokyo)
Metrics: PSNR SSIM LPIPS iterations to reach PSNR threshold
Baselines: Uniform sampling Edge (Gai et al.) Quadtree-based sampling (Zhang et al.) Multinomial sampling
Builds on: Instant-NGP NerfAcc NeRF Edge-based ray sampling hard example mining
Closest papers by meaning
- Learning Large-scale Neural Fields via Context Pruned Meta-Learning (NeurIPS 2023 · similarity 0.86)
- In Search of a Data Transformation That Accelerates Neural Field Training (CVPR 2024 · similarity 0.85)
- Repeated Random Sampling for Minimizing the Time-to-Accuracy of Learning (ICLR 2024 · similarity 0.84)
- LEAPS: A discrete neural sampler via locally equivariant networks (ICML 2025 · similarity 0.84)
- How to Train Neural Field Representations: A Comprehensive Study and Benchmark (CVPR 2024 · similarity 0.84)
- Fast Training of Sinusoidal Neural Fields via Scaling Initialization (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.