Research map / 3D vision / Face and shape
FINER: Flexible Spectral-bias Tuning in Implicit NEural Representation by Variable-periodic Activation Functions
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 FINER, an implicit neural representation using a variable-periodic activation sin((|x|+1)x) whose frequency set is tuned by bias initialization range, and evaluates on 2D image fitting, SDF, and NeRF.
Problem
Existing implicit neural representations have a restricted supported frequency set, so they struggle to represent complex signals with multiple frequencies (spectral bias and capacity-convergence gap in SIREN).
Approach
Replace the periodic sin activation with a variable-periodic function sin((|x|+1)x) and tune the supported frequency set by initializing the bias vector over a wider uniform range U(-k,k), selecting different sub-functions along the activation's domain; the method is motivated geometrically and via neural tangent kernel analysis.
Key results
- On 2D image fitting, FINER achieves PSNR 40.76 dB, SSIM 0.9790, LPIPS 2.56e-3, outperforming SIREN (38.52 dB) and others.
- On SDF representation, FINER reaches average Chamfer 3.087e-6 and IoU 0.9790, the best among compared INRs.
- On NeRF over 8 Blender scenes, FINER achieves best PSNR in most scenes (e.g., Ficus 28.70 vs SIREN 27.26).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace the SIREN activation with FINER's variable-periodic activation and flexible bias initialization | SIREN · Natural images, 512×512 | PSNR: 38.52 -> 40.76 | Helped |
| Replace the SIREN activation with FINER's variable-periodic activation and flexible bias initialization | SIREN · Natural images, 512×512 | SSIM: 0.9724 -> 0.9790 | Helped |
| Replace the SIREN activation with FINER's variable-periodic activation and flexible bias initialization | SIREN · Natural images, 512×512 | LPIPS: 5.52e-3 -> 2.56e-3 | Helped |
| Use FINER instead of WIRE | SDF representation · SDF, 5-shape average | Chamfer: 3.252e-6 -> 3.087e-6 | Helped |
| Use FINER instead of WIRE | SDF representation · SDF, 5-shape average | IOU: 9.760e-1 -> 9.790e-1 | Helped |
| Use FINER instead of the strongest baseline in each scene | NeRF · NeRF Blender, Chair | PSNR: 33.31 -> 33.90 | Helped |
| Use FINER instead of the strongest baseline in each scene | NeRF · NeRF Blender, Drums | PSNR: 24.89 -> 24.90 | Helped |
| Use FINER instead of the strongest baseline in each scene | NeRF · NeRF Blender, Ficus | PSNR: 27.26 -> 28.70 | Helped |
| Use FINER instead of the strongest baseline in each scene | NeRF · NeRF Blender, Hotdog | PSNR: 32.85 -> 33.05 | Helped |
| Use FINER instead of the strongest baseline in each scene | NeRF · NeRF Blender, Lego | PSNR: 29.60 -> 30.04 | Helped |
| Use FINER instead of SIREN | NeRF · NeRF Blender, Materials | PSNR: 27.13 -> 27.05 | Hurt |
| Use FINER instead of the strongest baseline in each scene | NeRF · NeRF Blender, Mic | PSNR: 33.59 -> 33.96 | Helped |
Surprising findings
- Under standard bias initialization, FINER's supported frequency set is small and loses high-frequency details, showing bias initialization range is the key control knob.
Limitations
- Bias initialization range k is set manually per task (1/sqrt(2), 1, 1/sqrt(3)), and gradient descent cannot guarantee reaching the global optimum from poor bias initialization.
- Analysis is limited to a simple 1D single-hidden-layer NTK case.
Experimental setup
Datasets: natural image dataset (16 images, 512x512) public SDF shapes (Armadillo, Dragon, Lucy, Thai Statue, BeardedMan) Blender NeRF dataset (8 scenes)
Benchmarks: natural image dataset (16 images) SDF shapes Armadillo/Dragon/Lucy/Thai Statue/BeardedMan Blender NeRF scenes
Metrics: PSNR SSIM LPIPS Chamfer distance IoU
Baselines: PEMLP SIREN Gauss WIRE
Builds on: SIREN NeRF positional encoding WIRE
Closest papers by meaning
- Implicit Neural Representations and the Algebra of Complex Wavelets (ICLR 2024 · similarity 0.87)
- FreSh: Frequency Shifting for Accelerated Neural Representation Learning (ICLR 2025 · similarity 0.87)
- A sampling theory perspective on activations for implicit neural representations (ICML 2024 · similarity 0.87)
- Inductive Gradient Adjustment for Spectral Bias in Implicit Neural Representations (ICML 2025 · similarity 0.87)
- Improved Implicit Neural Representation with Fourier Reparameterized Training (CVPR 2024 · similarity 0.86)
- Grids Often Outperform Implicit Neural Representation at Compressing Dense Signals (NeurIPS 2025 · similarity 0.86)
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.