Research map / Image generation / Image restoration
GLIC: General Format Learned 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 GLIC, a learned image compression model that unifies any-channel images into one-channel 2D matrices and adds an adaptive attention residual block and an evenly grouped cross-channel context module for progressive preview decoding.
Problem
Learned image compression models only handle three-channel images, which limits their use on general image formats (grayscale, multi-channel medical, spectral data) and hampers progressive decoding.
Approach
Reshape any C-channel image into a single-channel matrix via tiling so the entropy model avoids inter-channel correlation effects, and combine it with an adaptive attention residual block (AARB) and evenly grouped cross-channel context (CCCT) for progressive preview decoding.
Key results
- GLIC reports better rate-distortion than prior learned and hand-crafted codecs on Kodak in PSNR and MS-SSIM, claimed as SOTA.
- On LUNA16 (multi-channel), GLIC reaches 41.43 PSNR at 1.05 BPP; on MURA (one-channel) 44.79 PSNR at 0.41 BPP.
- Replacing the multi-channel model with the one-channel scheme significantly improves RD performance in the ablation on Kodak.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Convert multi-channel images into a one-channel 2D matrix before entropy modeling (one-channel scheme) (ablation) | learned image compression model · Kodak | RD performance (PSNR vs bpp) | Helped |
| Replace RAB with adaptive attention residual block (AARB) in the one-channel model (ablation) | one-channel GLIC model · Kodak | RD performance | Helped |
| Replace channel-wise context with evenly grouped cross-channel context (CCCT) (ablation) | GLIC preview decoding (half chunks, g(4)) · Kodak | preview image RD performance | Helped |
| Replace channel-wise context with CCCT (ablation) | GLIC full (8-chunk) decoding · Kodak | complete image RD performance | No clear effect |
| Use CCCT with multi-channel (three-channel) model instead of one-channel (ablation) | three-channel model · Kodak | RD performance and decode consistency | Hurt |
| Train with only the lowest preview synthesizers g(1,2,3) versus g(>3) | progressive decoding synthesizers · Kodak | RD performance | Hurt |
| One-channel GLIC compared to existing learned models (Ballé, Cheng, He2022, Minnen) and traditional codecs | GLIC · Kodak | BPP-PSNR and BPP-MS-SSIM | Helped |
Surprising findings
- Pairing a three-channel model with the CCCT module causes the model to inconsistently fail to encode and decode the same image, which the authors attribute to strong channel correlation.
Limitations
- Reported RD gains on general-format data are shown only for single-setting results without comparison to baselines on LUNA16, MURA, or Bluetooth data.
- Preview decoding quality is only slightly below full decoding, and lower preview levels (g1-g3) show a large RD gap.
- Training used MSE only and a subset of ImageNet; hardware and training cost not stated.
Experimental setup
Datasets: ImageNet subset LUNA16 MURA Bluetooth I/Q data
Benchmarks: Kodak LUNA16 MURA Bluetooth I/Q CLIC image
Metrics: PSNR MS-SSIM (dB) BPP
Baselines: Ballé2018 Minnen2018 Cheng2020 He2022 MLIC+ CTC Lee Johnston2018 Liu2023 JPEG JPEG2000 WebP
Builds on: Minnen and Singh 2020 channel context model He et al. 2022 unevenly grouped channel context He et al. 2021 checkerboard context Cheng et al. 2020 residual attention block Attention in Attention network (Chen et al. 2021)
Closest papers by meaning
- Adaptive Learned Image Compression with Graph Neural Networks (CVPR 2026 · similarity 0.89)
- Learned Image Compression With Discretized Gaussian Mixture Likelihoods and Attention Modules (CVPR 2020 · similarity 0.89)
- How To Exploit the Transferability of Learned Image Compression to Conventional Codecs (CVPR 2021 · similarity 0.88)
- Linear Attention Modeling for Learned Image Compression (CVPR 2025 · similarity 0.88)
- Learned Image Compression with Dictionary-based Entropy Model (CVPR 2025 · similarity 0.87)
- Good, Cheap, and Fast: Overfitted Image Compression with Wasserstein Distortion (CVPR 2025 · similarity 0.87)
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.