Research map / Image generation / Image restoration
Once-for-All: Controllable Generative Image Compression with Dynamic Granularity Adaptation
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.
Control-GIC is a VQGAN-based generative image compression model that controls bitrate at inference via granularity ratios over entropy-ranked patches, using one set of weights.
Problem
Generative image compression methods need separate models per bitrate, and existing variable-rate or progressive methods support only limited bitrate ranges or use MSE-oriented objectives, so flexible fine-grained rate adaptation with perceptual quality is lacking.
Approach
Allocate multi-granularity VQ codes to image patches by their spatial entropy, so the number of VQ indices (and bitrate) is tuned at inference by granularity ratios; a non-parametric statistical Huffman coder losslessly codes indices, and a probabilistic conditional decoder reinjects multi-granularity features.
Key results
- Control-GIC encodes and decodes fastest among compared methods: about 7x faster than MS-ILLM and 4x faster than MRIC in encoding, per the paper.
- Control-GIC achieves BD-rate saving close to MRIC and MS-ILLM on Kodak (LPIPS, VVC anchor) while using a single model trained for 0.6M steps.
- On extremely low bitrate (<0.05 bpp), Control-GIC achieves LPIPS 0.115 vs 0.136 for Mao et al. on Kodak and 0.086 vs 0.112 on CLIC2020.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use statistical entropy coding instead of Huffman coding with uniform frequency (ablation) | Control-GIC · Kodak, granularity ratio 100%, 0, 0 | bit cost: 0.625 -> 0.594; 5.0% bit saving | Helped |
| Use statistical entropy coding instead of Huffman coding with uniform frequency (ablation) | Control-GIC · Kodak, granularity ratio 60%, 30%, 10% | bit cost: 0.445 -> 0.432; 2.9% | Helped |
| Use statistical entropy coding instead of Huffman coding with uniform frequency (ablation) | Control-GIC · Kodak, granularity ratio 20%, 50%, 30% | bit cost: 0.235 -> 0.234; 0.4% | Helped |
| Add the medium-grained condition (med) to the decoder (ablation) | Control-GIC probabilistic conditional decoder · DIV2K | R-D performance | Helped |
| Add the fine-grained condition (fin) to the decoder (ablation) | Control-GIC probabilistic conditional decoder · DIV2K | R-D performance | Helped |
| Add both medium- and fine-grained conditions to the decoder (ablation) | Control-GIC probabilistic conditional decoder · DIV2K | DISTS | Helped |
| Remove both medium- and fine-grained conditions from the decoder (ablation) | Control-GIC probabilistic conditional decoder · DIV2K | R-D performance | Hurt |
| Reduce the medium-granularity ratio so the fine-granularity ratio increases (ablation) | Control-GIC · Kodak | bpp: 0.4171 -> 0.4172 | Mixed |
| Increase the fine- and medium-granularity ratios (ablation) | Control-GIC · Kodak | LPIPS | Helped |
Surprising findings
- Statistical entropy coding saves only 0.4% bits at the coarsest-heavy ratio, showing gains concentrate at high bpp.
Limitations
- Small faces and fine details are hard to reconstruct at coarse granularity ratios, as noted in the appendix.
- Actual bpp deviates from theoretical bpp by up to 0.05, requiring manual ratio adjustment.
- Training used a fixed granularity ratio (50%, 40%, 10%), with the effect of other training ratios not explored.
Experimental setup
Datasets: OpenImages Kodak DIV2K CLIC2020
Benchmarks: Kodak DIV2K CLIC2020
Metrics: LPIPS DISTS PSNR FID KID NIQE BD-rate encoding time decoding time bpp
Baselines: HiFiC MRIC MS-ILLM CDC SCR CTC M&S (Hyperprior) BPG VVC (VTM10.0) Mao et al. VQGAN compression
Builds on: VQGAN MoVQ HiFiC
Closest papers by meaning
- Generative Latent Coding for Ultra-Low Bitrate Image Compression (CVPR 2024 · similarity 0.87)
- HyperVQ: Enabling Hyperprior Entropy Modeling for VQ-Based Generative Image Compression (NeurIPS 2026 · similarity 0.87)
- Good, Cheap, and Fast: Overfitted Image Compression with Wasserstein Distortion (CVPR 2025 · similarity 0.86)
- Thinking in Granularity: Dynamic Quantization for Image Super-Resolution by Intriguing Multi-Granularity Clues (AAAI 2025 · similarity 0.86)
- Decouple Distortion from Perception: Region Adaptive Diffusion for Extreme-low Bitrate Perception Image Compression (CVPR 2025 · similarity 0.86)
- Towards image compression with perfect realism at ultra-low bitrates (ICLR 2024 · 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.