Research map / Probabilistic and scientific modeling / Uncertainty and Gaussian processes
RECOMBINER: Robust and Enhanced Compression with Bayesian 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.
Proposes RECOMBINER, an improved COMBINER INR-based compression method using linear weight reparameterization, learned positional encodings, and hierarchical Bayesian patch modeling with adaptive beta, evaluated on image, audio, video, and protein structure data.
Problem
Implicit neural representation (INR)-based compression with Bayesian relative entropy coding (COMBINER) is limited by factorized posteriors, difficulty overfitting small INRs at low bitrates, and sensitivity to modeling choices and variational initialization on high-resolution data.
Approach
Replace factorized Gaussian weight posteriors with a learned linear reparameterization w = h A (full-covariance-like but factorized to code), add learnable upsampled positional encodings as extra INR input, split high-resolution data into patches with a three-level hierarchical Gaussian prior, and adapt beta explicitly to a bit budget.
Key results
- On CIFAR-10 at low bitrates RECOMBINER outperforms even VAE-based codecs, and on Kodak it outperforms most INR-based codecs and is competitive with VC-INR while still trailing VAE codecs.
- Permutation across patches gives up to 1.5 dB gain at higher bitrates on Kodak and 1.00 dB on kodim23 at 0.074 bpp.
- Learnable positional encodings give about 2 dB gain at low bitrates on Kodak; linear reparameterization gives up to 4 dB gain on CIFAR-10 at high bitrates.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add linear weight reparameterization (ablation) | RECOMBINER · CIFAR-10 | PSNR: up to 4dB gain | Helped |
| add linear weight reparameterization (ablation) | RECOMBINER · Kodak | PSNR: over 0.5 dB gain | Helped |
| add learnable positional encodings (ablation) | RECOMBINER · CIFAR-10 below 2 bpp | PSNR: up to 0.5 dB gain | Helped |
| add learnable positional encodings (ablation) | RECOMBINER · Kodak at low bitrates | PSNR: 2 dB | Helped |
| add learnable positional encodings (ablation) | RECOMBINER · Kodak at high bitrates | PSNR: 1 dB | Helped |
| add hierarchical model without positional encodings (ablation) | RECOMBINER · CIFAR-10 and Kodak | PSNR | Hurt |
| add learnable positional encodings to the hierarchical model (ablation) | RECOMBINER · CIFAR-10 and Kodak | PSNR: 0.5 dB gain | Helped |
| add random permutation across patches (ablation) | RECOMBINER · CIFAR-10 and Kodak at low bitrates | PSNR: 0.5 dB | Helped |
| add random permutation across patches (ablation) | RECOMBINER · CIFAR-10 and Kodak at higher bitrates | PSNR: more than 1.5 dB | Helped |
| add random permutation across patches (ablation) | RECOMBINER · Kodak kodim23 at 0.074 bpp | PSNR: 1.00 dB gain | Helped |
| use RECOMBINER instead of COMBINER | INR image compression · CIFAR-10 | rate-distortion performance | Helped |
| use RECOMBINER instead of VAE-based codecs | low-resolution image compression · CIFAR-10 at low bitrates | rate-distortion performance | Helped |
Surprising findings
- The hierarchical model without positional encodings can degrade performance, attributed to optimization difficulty.
- Encoding time remains long, similar to COMBINER, despite fast decoding.
Limitations
- Encoding time complexity is very high, similar to COMBINER.
- Patch-based compression produces block artifacts.
- Approximate A* coding samples affect performance and require fine-tuning.
Experimental setup
Datasets: CIFAR-10 Kodak DIV2K LibriSpeech UCF-101 AlphaFold DB (S. cerevisiae proteome)
Benchmarks: CIFAR-10 Kodak LibriSpeech UCF-101 AlphaFold DB yeast proteome
Metrics: PSNR bits per pixel kbps RMSD bits per atom
Baselines: COMBINER VC-INR COIN++ COIN MSCN JPEG2000 BPG VTM H.264 H.265 MP3 Foldcomp
Builds on: COMBINER COIN++ SIREN relative entropy coding
Closest papers by meaning
- SINR: Sparsity Driven Compressed Implicit Neural Representations (CVPR 2025 · similarity 0.88)
- Diversify, Contextualize, and Adapt: Efficient Entropy Modeling for Neural Image Codec (NeurIPS 2024 · similarity 0.85)
- Robustly overfitting latents for flexible neural image compression (NeurIPS 2024 · similarity 0.85)
- MCNC: Manifold-Constrained Reparameterization for Neural Compression (ICLR 2025 · similarity 0.84)
- Learning Transferable Features for Implicit Neural Representations (NeurIPS 2024 · similarity 0.84)
- Grids Often Outperform Implicit Neural Representation at Compressing Dense Signals (NeurIPS 2025 · similarity 0.84)
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.