Spectral Prefiltering of Neural Fields

2025-10-09

AI Summary

Neural fields struggle to efficiently represent multiscale visual signals due to their inherent support for only a single fixed-resolution reconstruction. This work proposes a frequency-domain prefiltering framework enabling continuous-scale signal reconstruction via a single forward pass. Its core innovation is closed-form Fourier feature modulation: the frequency response of parameterized filters—such as Gaussian, Box, or Lanczos—is analytically embedded into Fourier features without altering the network architecture. Combined with single-sample Monte Carlo estimation, the method supports end-to-end training while preserving efficient inference and delivering high-fidelity prefiltering. Experiments demonstrate substantial improvements over existing neural field filtering approaches in training and inference speed, quantitative metrics (PSNR/SSIM), and perceptual quality.

Technology Category

• Computer Vision: Multi-modal Vision
• Machine Learning: Multimodal Learning
• Cognitive Modeling & Cognitive Systems: Neural Spike Coding

Application Category

• Search and Retrieval-Augmented AI: Efficiency and scalability of Web search engines
• User Modeling, Personalization and Recommendation: On-Device user modeling, personalization, and recommendation
• Web Mining and Content Analysis: Large pretrained models with web data

Abstract

Neural fields excel at representing continuous visual signals but typically operate at a single, fixed resolution. We present a simple yet powerful method to optimize neural fields that can be prefiltered in a single forward pass. Key innovations and features include: (1) We perform convolutional filtering in the input domain by analytically scaling Fourier feature embeddings with the filter's frequency response. (2) This closed-form modulation generalizes beyond Gaussian filtering and supports other parametric filters (Box and Lanczos) that are unseen at training time. (3) We train the neural field using single-sample Monte Carlo estimates of the filtered signal. Our method is fast during both training and inference, and imposes no additional constraints on the network architecture. We show quantitative and qualitative improvements over existing methods for neural-field filtering.

Problem

• Enables neural fields to operate at multiple resolutions efficiently
• Performs analytical prefiltering using Fourier feature modulation
• Supports generalization to unseen filter types during inference

Innovation

• Prefiltering neural fields via Fourier feature scaling
• Generalizing filtering with parametric unseen filters
• Training using single-sample Monte Carlo estimates