Modern image compression methods are typically optimized for the rate--distortion--perception trade-off, whereas their robustness to bit-level corruption is rarely examined. We show that diffusion-based compressors built on the Reverse Channel Coding (RCC) paradigm are substantially more robust to bit flips than classical and learned codecs. We further introduce a more robust variant of Turbo-DDCM that significantly improves robustness while only minimally affecting the rate--distortion--perception trade-off. Our findings suggest that RCC-based compression can yield more resilient compressed representations, potentially reducing reliance on error-correcting codes in highly noisy environments.
Figures & tables
Figure 1: Image Compression Methods Robustness to Bit-Flip Errors: The figure shows reconstructed images from the Kodak24 dataset after transmission through a noisy channel with bit-flip probabilities of 10−4 (first row) and 10−3 (second row). S.C. denotes StableCodec, and T-DDCM denotes Turbo-DDCM. At a bit-flip probability of 10−4 , RCC-based methods (the last three) perform well, whereas the others degrade significantly. At 10−3 , our proposed method, Robust Turbo-DDCM, is the only approach that maintains good reconstruction quality. All methods operate at bit-per-pixel rate, yielding high-fidelity reconstructions in a noise-free channel.
Figure 2: Quantitative Results: We compare the bit-flip robustness of several image compression methods, including non-neural, trained neural, and diffusion-based RCC methods, by evaluating distortion (PSNR or LPIPS) and perception (FID) across multiple BER values. Overall, RCC-based methods outperform the others, with Robust Turbo-DDCM demonstrating particularly strong robustness. The dashed vertical line in each subplot indicates the encoder–decoder distortion bound imposed by SD 2.1, obtained by passing clean images through the encoder–decoder without compression. Since all compared RCC-based methods rely on SD 2.1, they are all subject to this distortion bound.
Figure 3: Rate-Distortion-Perception Tradeoff: Robust Turbo-DDCM increases the bits per pixel compared to Turbo-DDCM. While it significantly improves robustness, it exhibits inferior rate–distortion–perception performance.
Figure 4: Qualitative Results: The presented reconstructions are based on images from the Kodak24 ( 512×512 ) dataset under a BER of 10−4 . While other methods produce reconstructions that aren’t resemble the original image, RCC-based methods maintains high fidelity.
Figure 5: Qualitative Results: The presented reconstructions are based on images from the Kodak24 ( 512×512 ) dataset under a BER of 10−3 . While other methods produce reconstructions that no longer resemble the original image, our method maintains high visual fidelity.
Appendix figures & tables2 assets
Supplementary material from the paper’s appendix.
Appendix
Dataset
Method
PSNR ↑
LPIPS ↓
FID ↓
% Corrupted Files ↓
Kodak24
JPEG
9.15 ± 2.54
0.75 ± 0.17
157.98 ± 24.56
0.83
BPG
10.43 ± 2.00
0.75 ± 0.11
120.03 ± 4.26
2.08
ILLM
17.43 ± 5.45
0.42 ± 0.22
75.05 ± 5.25
7.08
StableCodec
19.12 ± 4.81
0.38 ± 0.28
57.30 ± 6.82
0.42
DiffC
20.92 ± 4.18
0.29 ± 0.18
33.95 ± 1.74
0.87
DDCM
20.52 ± 3.70
0.30 ± 0.18
40.31 ± 6.17
0.00
Appendix
Table 2: Exact aggregated results for all methods at BER 10−4 . The table reports bit-flip robustness on Kodak24 and DIV2K using PSNR, LPIPS, FID, and the percentage of corrupted output files. Values are mean ± standard deviation, except corrupted files, which are reported as percentages. Bold indicates the best result for each metric and dataset.
Dataset
Method
PSNR ↑
LPIPS ↓
FID ↓
% Corrupted Files ↓
Kodak24
JPEG
7.84 ± 2.14
0.98 ± 0.08
420.95 ± 38.97
20.83
BPG
9.33 ± 1.56
0.84 ± 0.08
140.70 ± 6.28
15.83
ILLM
10.12 ± 2.78
0.75 ± 0.10
169.38 ± 9.16
61.67
StableCodec
13.18 ± 3.62
0.74 ± 0.21
129.21 ± 15.79
10.83
DiffC
14.85 ± 3.00
0.62 ± 0.15
57.94 ± 3.53
8.26
DDCM
16.74 ± 2.69
0.52 ± 0.16
76.57 ± 7.09
0.00
Appendix
Table 3: Exact aggregated results for all methods at BER 10−3 . The table reports bit-flip robustness on Kodak24 and DIV2K using PSNR, LPIPS, FID, and the percentage of corrupted output files. Values are mean ± standard deviation, except corrupted files, which are reported as percentages. Bold indicates the best result for each metric and dataset.
Department of Information Engineering Shanghai University of Electric Power Shanghai, China · Department of Computer Science and Communication Engineering Waseda University Tokyo, Japan · Faculty of Engineering Institute of Science Tokyo Tokyo, Japan