Research map / 3D vision / Face and shape

Enhancing Implicit Neural Representations via Symmetric Power Transformation

AAAI 2025 · 3D vision › Face and shape · approach: Implicit neural representation design · 3 citations (OpenAlex)

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 symmetric power transformation, a reversible, storage-free data transformation that maps signals into a defined range with approximately zero skewness to improve SIREN and FINER implicit neural representation training.

Problem

Implicit neural representations (INRs) train slowly and reach limited fitting quality; prior data transformations like random pixel permutation are irreversible, break continuity, and need extra storage.

Approach

Range-Defined Symmetric Hypothesis: INR fitting improves when signals are bounded to the activation range and have near-zero skewness. A power transformation normalized to [0,1] with coefficient beta (chosen so the median maps to 0.5), plus deviation-aware calibration and an adaptive soft boundary, realizes this reversibly with negligible overhead.

Key results

• On DIV2K at 5k iterations with SIREN backbone, sym-power reaches 36.75 dB PSNR vs 36.18 dB for scale baseline and 33.60 dB for 0-1 normalization.
• FINER + sym-power reaches 39.03 dB PSNR on DIV2K and 37.95 dB on Kodak at 5k iterations, state of the art among compared backbones.
• On 2D text fitting, FINER + sym-power improves PSNR from 27.56 to 34.30 dB at 0.5k iterations.

Measured effects reported in the paper

Surprising findings

• RPP, a prior data transformation, underperforms the plain baseline by roughly 15 dB under standard training, though it helps with very long training.
• Inverse transformation (flipping skewness sign) performs nearly as well as sym-power on DIV2K at 5k iterations (36.37 vs 36.75 dB).

Limitations

• Theory is heuristic; skewness-reduction analysis uses a log-normal model and the authors acknowledge lacking rigorous guarantees.
• Soft boundary gives only small average PSNR gains (0.05-0.1 dB) because few samples have boundary issues.
• Hyperparameters (xi, tau, kappa) are set manually and the gains on natural images are modest.

Experimental setup

Datasets: DIV2K Kodak text image dataset LibriSpeech UVG (ShakeNDry) simulated normal-distribution signals

Benchmarks: DIV2K Kodak LibriSpeech test.clean UVG ShakeNDry

Metrics: PSNR SSIM MSE STOI SI-SNR skewness range

Baselines: 0-1 normalization z-score standardization linear scaling inverse transform RPP Box-Cox PEMLP Gauss WIRE

Builds on: SIREN FINER Random pixel permutation (RPP) Box-Cox transformation DINER

Closest papers by meaning

• Grids Often Outperform Implicit Neural Representation at Compressing Dense Signals (NeurIPS 2025 · similarity 0.87)
• Towards Lossless Implicit Neural Representation via Bit Plane Decomposition (CVPR 2025 · similarity 0.87)
• Boosting Neural Representations for Videos with a Conditional Decoder (CVPR 2024 · similarity 0.86)
• Learning Input Encodings for Kernel-Optimal Implicit Neural Representations (ICML 2025 · similarity 0.86)
• VI^3NR: Variance Informed Initialization for Implicit Neural Representations (CVPR 2025 · similarity 0.85)
• Few-shot Implicit Function Generation via Equivariance (CVPR 2025 · similarity 0.85)