Subtractive Modulative Network (SMN) with Learnable Periodic Activations

Official implementation of “Subtractive Modulative Network with Learnable Periodic Activations” (IEEE ICASSP 2026).

Resources

• Paper (PDF): docs/SMN_2601.pdf
• Project Page: https://inrainbws.github.io/smn/
• Supplementary Materials: https://inrainbws.github.io/smn/
• ICASSP 2026 Accepted Papers (search by title): https://cmsworkshops.com/ICASSP2026/papers/accepted_papers.php

Abstract

We propose the Subtractive Modulative Network (SMN), a parameter-efficient Implicit Neural Representation (INR) architecture inspired by subtractive synthesis. SMN is structured as a signal-processing pipeline with (i) an Oscillator—a learnable periodic activation layer that generates a multi-frequency basis—and (ii) Filters—modulative mask modules that generate high-order harmonics. We provide theoretical analysis and empirical validation, achieving 40+ dB PSNR on two image datasets and showing consistent advantages on 3D NeRF novel view synthesis.

Quick Start

2D Image Representation (Kodak / DIV2K)

python inr_base/app/train.py

3D NeRF Synthesis (e.g., Lego)

python nerf/run_nerf.py --config configs/nerf_lego.txt