Research map / Probabilistic and scientific modeling / Variational and generative sampling
Learned Lossless Image Compression based on Bit Plane Slicing
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 ArIB-BPS, which splits 8-bit images into bit planes (ordered MSP to LSP) for autoregressive initial bits lossless compression, plus a dimension-tailored autoregressive model over plane, space, and color dimensions.
Problem
Lossless image compression with latent variable models and subimage autoregression (ArIB) suffers from uneven use of latent variables across spatial subimages and posterior collapse as the autoregressive prior grows.
Approach
Split images along the bit-plane dimension so that significant planes (where latent variables matter most) are coded with latents and insignificant planes serve as initial bits, arrange subimages by decreasing latent importance, and tailor autoregression per dimension (U-Net for planes, checkerboard small net for space, non-network Bernoulli for color).
Key results
- ArIB-BPS achieves 3.06 BPD on CIFAR10 dataset compression vs. 3.12 for iFlow and 3.16 for SHVC, and 3.07 single-image BPD on CIFAR10.
- On ImageNet64 dataset compression, ArIB-BPS reaches 3.63 BPD vs. 3.70 for iFlow, with 107.80 ms/sample (small variant 43.65 ms/sample).
- On DIV2K single-image compression (model trained on ImageNet64), ArIB-BPS achieves 2.59 BPD, the best among compared methods.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use ArIB-BPS instead of iFlow | CIFAR10 · CIFAR10 dataset compression | BPD: 3.118 -> 3.048 | Helped |
| use ArIB-BPS instead of iFlow | ImageNet32 · ImageNet32 dataset compression | BPD: 3.873 -> 3.911 | Hurt |
| use ArIB-BPS instead of iFlow | ImageNet64 · ImageNet64 dataset compression | BPD: 3.703 -> 3.633 | Helped |
| use ArIB-BPS instead of iFlow | CIFAR10 · CIFAR10 single-image compression | BPD: 37.398 -> 3.070 | Helped |
| use ArIB-BPS instead of iFlow | ImageNet32 · ImageNet32 single-image compression | BPD: 38.273 -> 3.918 | Helped |
| use ArIB-BPS instead of iFlow | ImageNet64 · ImageNet64 single-image compression | BPD: 38.123 -> 3.641 | Helped |
| add plane autoregression (ablation) | ArIB-BPS · CIFAR10 | theoretical BPD: 3.458 -> 3.052 | Helped |
| add space autoregression (ablation) | ArIB-BPS · CIFAR10 | theoretical BPD: 3.470 -> 3.052 | Helped |
| add color autoregression (ablation) | ArIB-BPS · CIFAR10 | theoretical BPD: 3.522 -> 3.052 | Helped |
| replace dimension-tailored space autoregression with the Φp architecture (ablation) | ArIB-BPS · CIFAR10 | theoretical BPD: 3.052 -> 3.045 | Helped |
| replace dimension-tailored space and color autoregression with the Φp architecture (ablation) | ArIB-BPS · CIFAR10 | theoretical BPD: 3.052 -> 3.015 | Helped |
| use the small plane-space model (ablation) | ArIB-BPS · CIFAR10 | theoretical BPD: 3.052 -> 3.125 | Hurt |
Surprising findings
- Ablation shows the Bernoulli-based color autoregression adds only negligible inference overhead, and adding plane autoregression with the heavy U-Net is needed for the main gains.
- Plane autoregression with a heavy U-Net mainly helps insignificant planes, whereas the dimension-tailored design performs better on significant planes with lower complexity.
Limitations
- Complexity grows with bit depth due to interdependency among bit planes; only 8-bit images are studied.
- Single-image results for SHVC and SHVC-ArIB are only available as lower bounds.
Experimental setup
Datasets: CIFAR10 ImageNet32 ImageNet64
Benchmarks: CIFAR10 ImageNet32 ImageNet64 DIV2K
Metrics: bits per dimension (BPD) inference time (ms/sample)
Baselines: SHVC-ArIB SHVC LBB iFlow iVPF IDF++ Bit-Swap Hilloc L3C LC-FDNet PNG FLIF
Builds on: ArIB (SHVC-ArIB) LBB iFlow PixelCNN++ LC-FDNet
Closest papers by meaning
- Joint Autoregressive and Hierarchical Priors for Learned Image Compression (NeurIPS 2018 · similarity 0.88)
- Towards Lossless Implicit Neural Representation via Bit Plane Decomposition (CVPR 2025 · similarity 0.85)
- Compression of Structured Data with Autoencoders: Provable Benefit of Nonlinearities and Depth (ICML 2024 · similarity 0.84)
- Distilled Decoding 1: One-step Sampling of Image Auto-regressive Models with Flow Matching (ICLR 2025 · similarity 0.83)
- CAT-3DGS: A Context-Adaptive Triplane Approach to Rate-Distortion-Optimized 3DGS Compression (ICLR 2025 · similarity 0.83)
- Robustly overfitting latents for flexible neural image compression (NeurIPS 2024 · similarity 0.83)
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.