Organizations: School of Electrical Engineering and Computer Science, KTH Royal Institute of Technology, Sweden · School of Computer and Communication Sciences, EPFL, Switzerland
We propose the Subtractive Modulative Network (SMN), a novel, parameter-efficient Implicit Neural Representation (INR) architecture inspired by classical subtractive synthesis. The SMN is designed as a principled signal processing pipeline, featuring a learnable periodic activation layer (Oscillator) that generates a multi-frequency basis, and a series of modulative mask modules (Filters) that actively generate high-order harmonics. We provide both theoretical analysis and empirical validation for our design. Our SMN achieves a PSNR of 40+ dB on two image datasets, comparing favorably against state-of-the-art methods in terms of both reconstruction accuracy and parameter efficiency. Furthermore, consistent advantage is observed on the challenging 3D NeRF novel view synthesis task. Supplementary materials are available at https://inrainbws.github.io/smn/.
Figures & tables
Figure 1: The end-to-end architecture of our SMN, illustrating the Oscillator, the multi-stage Filter with its main and masking pathways, and the final Amplifier (Self-Mask) stage.
Method
Kodak(dB)
DIV2K(dB)
Parameters
MLP
28.63
30.21
272415
Gauss [ 11 ]
37.90
38.34
272703
SIREN [ 13 ]
33.65
33.73
272703
WIRE [ 12 ]
40.24
38.90
265523
RINR [ 7 ]
32.96
34.03
289716
SMN (Ours)
41.40
42.53
264216
Table 1: Quantitative comparison on 2D image representation. Our SMN achieves the highest fidelity on both datasets.
Figure 2: Visual comparison of reconstruction quality for different INR methods on a Kodak image. Our method preserves fine textures and edges most faithfully. Best viewed on screen when zooming in.
Method
PSNR (dB)
Parameters
PE+Gauss [ 11 ]
32.00
287749
PE+SIREN [ 13 ]
29.06
287749
PE+WIRE [ 12 ]
25.14
283479
PE+RINR [ 7 ]
26.84
314703
PE+MLP [ 8 ]
26.66
290370
PE+SMN (Ours)
32.98
287749
Table 2: Average PSNR (dB) on the eight scenes of the NeRF synthetic dataset at 400 × 400 resolution.
Figure 3: Qualitative comparison of view synthesis on the Lego NeRF dataset. Our method reconstructs fine geometric details more faithfully and reduces common artifacts such as floater noise and blurriness. Best viewed on screen when zooming in.
Variant ID
Description
PSNR (dB)
Variant 1
Fixed Amplitudes
35.08
Variant 2
K=1
42.87
Variant 3
K=2
43.09
Final Design
K=3
43.68
Table 3: Ablation study of the Learnable Sine Layer’s activation function. Performance is measured by average PSNR (dB) on the Kodak dataset.