Research map / Image generation / Image restoration
Decouple Distortion from Perception: Region Adaptive Diffusion for Extreme-low Bitrate Perception 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 MRIDC, a VQ-latent diffusion image codec that masks latent tokens according to RoI maps (training-free) and uses a bidirectional mask-prediction transformer plus ControlNet to decode extreme-low-bitrate images with better regional fidelity.
Problem
Generative image compression at extremely low bitrates (~0.01-0.05 bpp) treats images uniformly, failing to balance global perceptual quality with local distortion in regions of interest and to allocate coding resources to complex areas.
Approach
Encode VQ latent tokens and mask non-RoI tokens according to a pixel-domain RoI map (Map-guided Latent Masking, training-free), then reconstruct masked tokens with a bidirectional mask prediction transformer (BPCG) whose output controls a diffusion decoder via ControlNet, starting denoising from a global semantic latent zh.
Key results
- At 0.01 and 0.04 bpp base rates, MRIDC matches PerCo's perceptual quality while giving superior distortion metrics even without RoI enabled.
- Removing the global semantic latent zh raises FID from 43.9 to 55.1 on MSCOCO411 ablation (w/ zh vs w/o zh).
- Encode/decode time on Kodak: MRIDC (30% mask) 0.083s/1.35s vs PerCo 0.04s/0.98s and PICS 221.10s/42.80s on a single A800.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add global semantic information zh (ablation) | MRIDC · MSCOCO411 | PSNR: 18.0 -> 20.3; RoI: 17.0 -> 18.9 | Helped |
| add global semantic information zh (ablation) | MRIDC · MSCOCO411 | FID: 55.1 -> 43.9 | Helped |
| add MPT latent prediction (ablation) | MRIDC · MSCOCO411 | PSNR: Image: 16.4 -> 19.3; RoI: 15.7 -> 17.9 | Helped |
| add MPT latent prediction (ablation) | MRIDC · MSCOCO411 | FID: 73.2 -> 49.4 | Helped |
| increase masking ratio from 20% to 40% (ablation) | MRIDC with Bbox RoI map · MSCOCO411, Low rate | FID: 48.2 -> 54.8 | Hurt |
| increase masking ratio from 20% to 40% (ablation) | MRIDC with Edge RoI map · MSCOCO411, Low rate | FID: 52.1 -> 66.8 | Hurt |
| increase masking ratio from 20% to 40% (ablation) | MRIDC with Rand RoI map · MSCOCO411, Low rate | FID: 51.0 -> 62.8 | Hurt |
| use Bbox instead of Edge RoI map at 30% masking (ablation) | MRIDC · MSCOCO411, Low rate | FID: 58.6 -> 50.7 | Helped |
| use Bbox instead of Edge RoI map at 30% masking (ablation) | MRIDC · MSCOCO411, High rate | FID: 52.4 -> 46.1 | Helped |
Limitations
- Bitrate is bounded by the size of the predefined latent grid, limiting variable-rate flexibility.
- Encoding and decoding are slower than PerCo, though faster than PICS.
- RoI maps must be provided by user annotation, edge detection, or bounding boxes.
Experimental setup
Datasets: OpenImages-v6
Benchmarks: MSCOCO30K MSCOCO411 Kodak
Metrics: bpp PSNR MS-SSIM LPIPS DISTS FID KID
Baselines: VTM ELIC HiFiC MS-ILLM MRIC PICS PerCo
Builds on: PerCo ControlNet Stable Diffusion v2.1 VQGAN RDP theory
Closest papers by meaning
- Controllable Distortion-Perception Tradeoff Through Latent Diffusion for Neural Image Compression (AAAI 2025 · similarity 0.92)
- Towards image compression with perfect realism at ultra-low bitrates (ICLR 2024 · similarity 0.92)
- Correcting Diffusion-Based Perceptual Image Compression with Privileged End-to-End Decoder (ICML 2024 · similarity 0.90)
- Laplacian-guided Entropy Model in Neural Codec with Blur-dissipated Synthesis (CVPR 2024 · similarity 0.89)
- Ultra Lowrate Image Compression with Semantic Residual Coding and Compression-aware Diffusion (ICML 2025 · similarity 0.89)
- Bridging the Gap between Gaussian Diffusion Models and Universal Quantization for Image Compression (CVPR 2025 · similarity 0.89)
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.