Research map / Architectures and efficiency / Pruning and efficient tuning
CALLIC: Content Adaptive Learning for 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 CALLIC, a content-adaptive learned lossless image compressor that pretrains a lightweight masked gated ConvFormer and then fine-tunes low-rank incremental weights per test image under an MDL objective, using rate-guided progressive patch selection.
Problem
Learned lossless image compression models are amortized over large training sets, giving sub-optimal probability estimates for specific test images and slow per-image adaptation.
Approach
Connects the two-stage MDL principle with parameter-efficient transfer learning: a pretrained autoregressive model (MGCF) is adapted per image via low-rank (LoRA/Tucker) incremental weights that are merged after tuning and transmitted as a compact bitstream prompt, with RPFT progressively adding high-entropy patches to speed adaptation.
Key results
- CALLIC achieves 2.54 bpsp on Kodak, 1.74 on RS19, 2.74 on Histo24, 2.46 on DIV2K and 2.30 on CLIC.p, beating all listed codecs and learned methods.
- MGCF (575K parameters) reaches 2.77 bpsp on Kodak versus 2.78 for ArIB-BPS (146.6M parameters).
- RPFT reduces CALLIC encoding time on Kodak from 12.9s to 9.7s with unchanged bpsp (2.54).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add image-specific RPFT adaptation to MGCF | MGCF · Kodak | bpsp: 2.77 -> 2.54 | Helped |
| add image-specific RPFT adaptation to MGCF | MGCF · RS19 | bpsp: 1.94 -> 1.74 | Helped |
| add image-specific RPFT adaptation to MGCF | MGCF · Histo24 | bpsp: 2.88 -> 2.74 | Helped |
| remove CCI (ablation) | MGCF · Kodak | encoding time: 1.4s -> 15.2s | Hurt |
| remove CCI (ablation) | MGCF · Kodak | decoding time: 1.6s -> 15.4s | Hurt |
| remove RPFT (ablation) | CALLIC · Kodak | encoding time: 9.7s -> 12.9s | Hurt |
| replace masked convolutional gating with masked neighborhood attention (ablation) | MGCF · Kodak | bpsp: 2.77 -> 2.78 | Hurt |
| reduce depth from 3 to 1 (ablation) | MGCF · Kodak | bpsp: 2.77 -> 2.84 | Hurt |
| increase depth from 3 to 5 (ablation) | MGCF · Kodak | bpsp: 2.77 -> 2.73 | Helped |
| reduce dimension from 128 to 64 (ablation) | MGCF · Kodak | bpsp: 2.77 -> 2.82 | Hurt |
| increase dimension from 128 to 192 (ablation) | MGCF · Kodak | bpsp: 2.77 -> 2.74 | Helped |
| increase dimension from 128 to 256 (ablation) | MGCF · Kodak | bpsp: 2.77 -> 2.72 | Helped |
Surprising findings
- Lightweight pretrained model with 575K parameters outperforms a 146.6M-parameter ArIB-BPS on most datasets.
- Starting adaptation from low-entropy patches ('increasing entropy') or random sampling is worse than the default high-entropy-first ordering.
Limitations
- Per-image adaptation increases encoding time (9.7s on Kodak vs 1.4s for MGCF alone).
- Gains over the pretrained MGCF are smaller on datasets close to the training distribution (DIV2K, CLIC.p).
Experimental setup
Datasets: DIV2K Flickr2K Kodak WHU-RS19 Histo24 CLIC2020 professional validation
Benchmarks: Kodak RS19 Histo24 DIV2K CLIC.p
Metrics: bits per sub-pixel (bpsp) encoding time decoding time parameter count
Baselines: PNG JPEG2000 FLIF JPEG-XL L3C RC iVPF LC-FDNet DLPR ArIB-BPS
Builds on: LoRA MetaFormer DLPR Neighborhood Attention Transformer
Closest papers by meaning
- Practical Full Resolution Learned Lossless Image Compression (CVPR 2019 · similarity 0.87)
- GLIC: General Format Learned Image Compression (AAAI 2025 · similarity 0.85)
- Large Language Models for Lossless Image Compression: Next-Pixel Prediction in Language Space is All You Need (NeurIPS 2025 · similarity 0.84)
- Once-for-All: Controllable Generative Image Compression with Dynamic Granularity Adaptation (ICLR 2025 · similarity 0.84)
- Linear Attention Modeling for Learned Image Compression (CVPR 2025 · similarity 0.84)
- MambaIC: State Space Models for High-Performance Learned Image Compression (CVPR 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.