Research map / Architectures and efficiency / RNNs and normalization
SINR: Sparsity Driven Compressed Implicit Neural Representations
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.
Compresses trained implicit neural representation weights by sparse-coding each weight vector over a randomly seeded Gaussian dictionary (via OMP), then applying quantization and Brotli coding, improving storage for images, occupancy fields, and NeRFs.
Problem
Storage-efficient compression of implicit neural representations (INRs) across modalities, where existing methods rely on quantization/entropy coding of the trained INR or learned latent transformations.
Approach
Observes that INR weight vectors are approximately Gaussian, so by the central limit theorem a random Gaussian dictionary regenerated from a seed can be used; a sparse code over this dictionary (found by L1 minimization / OMP) is stored instead of the weights, so the dictionary need not be learned or transmitted.
Key results
- On KODAK, reaching about 30 dB PSNR requires ~3.7 bpp for COIN and ~2.0 bpp for INRIC, while SINR reduces this to about 1.7 bpp with the same quantizer and entropy coder.
- For COIN++ on KODAK, SINR reaches ~24.2 dB PSNR at just under 1 bpp versus more than 1.5 bpp for COIN++.
- SINR achieves the smallest file size and highest IoU for every occupancy shape tested in the Stanford shape experiments.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply SINR compression to INR weights before quantization and entropy coding | INR image compression on KODAK · KODAK | bpp at average PSNR 30 dB: about 3.7 bpp (COIN) and 2.0 bpp (INRIC) -> approximately 1.7 bpp | Helped |
| Apply SINR compression to COIN++ latent modulations | COIN++ · KODAK | bpp at average PSNR approximately 24.2 dB: more than 1.5 bpp -> just under 1 bpp | Helped |
| Increase the number of hidden neurons in the INR | SINR image encoding · KODAK, experiment C1 | bpp at comparable PSNR | Helped |
| Increase the number of hidden layers while keeping the neuron count at 64 | SINR image encoding · KODAK, experiment C2 | bpp relative to INRIC | Mixed |
Surprising findings
- The dictionary atoms do not need to be learned or transmitted; a seeded random Gaussian matrix suffices to recover INR weights.
Limitations
- Choice of sparsity level s depends on neuron count and must be tuned incrementally.
- Pruning and L1 regularization alternatives degrade reconstruction quality, motivating the approach but indicating sensitivity.
- Tiny INRs with few neurons are hard to sparsify directly, requiring flattening of weight matrices.
Experimental setup
Datasets: KODAK Stanford shape dataset
Benchmarks: KODAK Stanford shape dataset NeRF
Metrics: PSNR bits per pixel file size IoU
Baselines: COIN COIN++ INRIC SHACIRA
Builds on: COIN COIN++ INRIC SHACIRA WIRE
Closest papers by meaning
- NVRC: Neural Video Representation Compression (NeurIPS 2024 · similarity 0.89)
- Grids Often Outperform Implicit Neural Representation at Compressing Dense Signals (NeurIPS 2025 · similarity 0.89)
- Implicit Representations via Operator Learning (ICML 2024 · similarity 0.89)
- RECOMBINER: Robust and Enhanced Compression with Bayesian Implicit Neural Representations (ICLR 2024 · similarity 0.88)
- Towards Lossless Implicit Neural Representation via Bit Plane Decomposition (CVPR 2025 · similarity 0.88)
- Learning Transferable Features for Implicit Neural Representations (NeurIPS 2024 · similarity 0.88)
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.