Research map / Image generation / Image restoration
Fitted Neural Lossless Image Compression
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 FNLIC, a lossless image codec that combines a per-image overfitted latent variable model with a dataset-pre-fitted lightweight checkerboard autoregressive prior, coordinated via independent fitting and a learned scale map.
Problem
Neural lossless image compression fitted to large datasets needs complex networks for intricate distributions, causing high decoding complexity; per-image overfitting adds parameter bitrate overhead.
Approach
Overfit a latent variable model per image for lossless compression, and pre-fit a lightweight autoregressive prior on a large dataset; coordinate them through independent fitting of the pre-fitter and an overfitted per-subimage scale map that adaptively transforms the prior distribution parameters.
Key results
- FNLIC achieves 2.88 BPD on Kodak, close to JPEG-XL (2.87) and comparable to neural methods like LC-FDNet (2.93) and ArIB-BPS (2.78).
- FNLIC decodes a 768x512 image in 13 ms on GPU and 221 ms on CPU, about 45x faster on GPU than LC-FDNet while matching its compression.
- Overfitting reduces Kodak BPD from 3.36 to 2.93 (about 14.3%), and pre-fitting further reduces it to 2.88.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FNLIC instead of JPEG-XL | lossless image compression · CLIC2024, DIV2K, Histo24, LoveDA24, Doc24 | BPD: 2.76 -> 2.68; 2.79 -> 2.75; 4.08 -> 3.86; 3.32 -> 3.30; 2.30 -> 2.22 | Helped |
| Use FNLIC instead of JPEG-XL | lossless image compression · Kodak | BPD: 2.87 -> 2.88 | Hurt |
| Use FNLIC instead of LC-FDNet | lossless image compression · Kodak | GPU inference time: 607 -> 13 | Helped |
| Add overfitting to the latent variable model (ablation) | FNLIC · Kodak | BPD: 3.36 -> 2.88 | Helped |
| Add pre-fitting (ablation) | overfitted latent variable model · Kodak | BPD: 2.93 -> 2.88 | Helped |
| Add independent fitting and adaptive prior transformation (ablation) | FNLIC · Kodak | BPD: 2.93 -> 2.88 | Helped |
| Use surrogate fitting and fixed prior scale instead of independent fitting and adaptive transformation (ablation) | FNLIC · Kodak | BPD: 2.88 -> 2.93 | Hurt |
| Remove adaptive prior transformation (ablation) | FNLIC with independently fitted pre-fitter · Kodak | BPD: 2.88 -> 2.91 | Hurt |
| Fix prior scale to 1 while using independent fitting (ablation) | FNLIC · Kodak | BPD: 2.88 -> 2.93 | Hurt |
Surprising findings
- Pre-fitting a surrogate latent variable model and fusing it without a scale map degraded performance versus the overfitter alone (2.93 vs 2.88 BPD with the proposed scheme).
Limitations
- High encoding complexity due to per-image overfitting (~140K iterations), unsuitable for real-time encoding.
- Network parameter bitrate is not optimized, hurting performance on small-resolution images.
- Overall compression ratio only slightly better or comparable to top neural and traditional codecs; trails DLPR and ArIB-BPS on compression.
Experimental setup
Datasets: Open Images Kodak CLIC2024 DIV2K ANHIR LoveDA Rumsey ICDAR'24 MapText
Benchmarks: Kodak CLIC2024 DIV2K Histo24 LoveDA24 Doc24
Metrics: bits per dimension decoding inference time (GPU and CPU)
Baselines: PNG JPEG-LS JPEG2000 WebP FLIF JPEG-XL L3C PILC LLICTI LC-FDNet DLPR ArIB-BPS
Builds on: C3 L3C FSAR Cool-Chic checkerboard autoregression
Closest papers by meaning
- Robustly overfitting latents for flexible neural image compression (NeurIPS 2024 · similarity 0.87)
- Neural Image Compression: Generalization, Robustness, and Spectral Biases (NeurIPS 2023 · similarity 0.87)
- Full Resolution Image Compression With Recurrent Neural Networks (CVPR 2017 · similarity 0.87)
- Neural Data-Dependent Transform for Learned Image Compression (CVPR 2022 · similarity 0.86)
- Lossless Compression with Probabilistic Circuits (ICLR 2022 · similarity 0.86)
- Distribution-Aware Tensor Decomposition for Compression of Convolutional Neural Networks (NeurIPS 2025 · similarity 0.85)
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.