Learned image compression (LIC) is trained almost exclusively on clean images: the only stochastic perturbation in the loop is an additive uniform proxy injected in the latent domain to keep the entropy model differentiable. Outside compression, by contrast, a long line of work---from Tikhonov regularization and contractive autoencoders through denoising score matching and modern diffusion models---has established that Gaussian noise at the input of a network with a clean target acts as a structured probe of the data manifold rather than as data augmentation. Building on this view, we introduce noise-regularized training for LIC: at every training step we add Gaussian noise to the encoder input while keeping the distortion target, the latent uniform proxy, the hyperprior, and the entropy coding pipeline unchanged. The recipe is architecture-agnostic and adds essentially no training or inference overhead. Across backbones and operating regimes it yields consistent BD-Rate gains: -2.60% on a reproduced ELIC and -0.85% / -1.05% / -2.38% on the Base/Medium/Large scales of a variable-rate DMCI codec, with DMCI-Large reaching -22.02% against VTM-22.0. A second-order expansion of the noisy RD loss accounts for these gains: because the noise enters upstream of the analysis transform, it converts the clean RD objective into a structured O(σ2) regularizer comprising an information-weighted encoder-Jacobian penalty (rate side) and a Frobenius contractive penalty on the end-to-end reconstruction map (distortion side). A finite-difference Lipschitz analysis shows corresponding sensitivity reductions of ~12% for the encoder and ~8% for the end-to-end codec, while an exact-JVP test further shows preferential contraction of high-information directions. Code and trained models are available at https://github.com/Googolxx/NoiseRegu-LIC.
Figures & tables
Model
Variant
Kodak
Tecnick
CLIC
Avg.
ELIC
baseline
0.00%
0.00%
0.00%
0.00%
w/ Noise Reg. (Ours)
−2.10%
−3.28%
−2.42%
−2.60%
DMCI
Base (baseline)
0.00%
0.00%
0.00%
0.00%
w/ Noise Reg. (Ours)
−0.90%
−0.85%
−0.80%
−0.85%
official release †
+2.91%
+5.60%
+5.45%
+4.66%
Medium (baseline)
0.00%
0.00%
0.00%
0.00%
Table 1: BD-Rate on Kodak, Tecnick, and CLIC Pro Valid. Each sub-group is anchored to its own baseline ( 0.00% ). All models reproduced under a unified training pipeline unless otherwise noted. Negative: bitrate savings at matched PSNR.
Initial noise scale σ0
0.125
0.25
0.50
0.75
BD-Rate on Kodak
+0.01%
−0.16%
−0.90%
−0.86%
Table 2: Sensitivity of noise-regularized DMCI-Base to the initial noise scale σ0 on Kodak. BD-Rate anchored to our reproduced DMCI-Base (clean training).
Method
kMACs/pixel
Param. (M)
Type
Kodak
Tecnick
CLIC
Avg.
TCM
1824
77
Single-Rate
−10.46%
−10.91%
−9.23%
−10.20%
MLIC++
1283
117
Single-Rate
−13.61%
−17.18%
−14.50%
−15.10%
LALIC
667
66
Single-Rate
−14.01%
−17.71%
−15.61%
−15.78%
DCAE
940
119
Single-Rate
−15.72%
−20.16%
−17.13%
−17.67%
HPCM-Base
919
69
Single-Rate
−14.67%
−17.37%
−15.26%
−15.77%
HPCM-Large
1261
90
Single-Rate
−18.70%
−21.44%
−19.43%
−19.86%
Table 3: Main RD comparison against VTM-22.0 in BD-Rate on Kodak, Tecnick, and CLIC. Negative values indicate bitrate savings at matched PSNR. Our noise-regularized variants are highlighted in bold; the best overall result is additionally bolded in the metric columns.
Figure 1: Local Lipschitz analysis on Kodak (DMCI-Base, δ=0.01 ). Top row, encoder L^ga (tests ΩR(1) ); bottom row, end-to-end codec L^f on the quantization-free map f=gs∘ga (tests ΩD(1)=∥Jf∥F2 ). Left panels (a,c): mean L^ vs. QP for the baseline and the noise-regularized model. Right panels (b,d): per-image distribution at QP =27 , estimated with n=64 random directions. Noise regularization contracts both Jacobians at every operating point ( 12.0% encoder, 8.4% codec on average; see Table 5 ).
QP
0
9
18
27
36
45
54
63
Weighted 1−Rw
39.15%
41.41%
46.56%
50.99%
46.08%
40.52%
38.05%
38.03%
Unweighted 1−Ru
29.75%
27.77%
31.13%
33.27%
32.28%
30.52%
30.13%
30.33%
Gap (pp)
+9.41
+13.64
+15.43
+17.72
+13.80
+10.00
+7.91
+7.70
Table 4: Encoder information-weighting test on DMCI-Large with exact JVPs and a shared noise-model Iϕ . Weighted contraction exceeds unweighted contraction at every QP, yielding a positive gap of 7.70 – 17.72 percentage points.
Probe scale δ
0.005
0.01
0.02
0.05
QP range
Encoder L^ga (tests ΩR(1) )
11.69%
12.00%
12.23%
12.24%
08.5% – 14.8%
End-to-end L^f (tests ΩD(1) )
08.21%
08.36%
08.45%
08.52%
07.4% – 10.0%
Table 5: Summary of relative reduction in local Lipschitz L^ on Kodak DMCI-Base (baseline → noise-regularized). Encoder column tests ΩR(1) ; end-to-end column tests ΩD(1) . Each cell averages Eq. ( 9 ) over 24 Kodak images, 8 QPs, and n=64 random directions. Full per-QP per- δ tables are in Appendix A.9 .
Dataset
PSNR
MS-SSIM
LPIPS
DISTS
Kodak
−1.63%
−1.99%
−2.13%
−2.84%
Tecnick
−0.63%
−0.77%
−1.12%
−1.94%
CLIC
−1.01%
−1.49%
−2.83%
−4.85%
Table 6: Perceptual DMCI-Large: BD-Rate of noise-regularized training against the corresponding clean model. Negative values indicate bitrate savings.
Dataset
Model
PSNR
MS-SSIM
LPIPS
DISTS
Kodak
Clean
−14.20%
−18.29%
−90.41%
−90.67%
Kodak
NoiseRegu
−15.75%
−20.04%
−90.66%
−91.01%
Tecnick
Clean
−17.10%
−14.62%
−87.34%
−88.29%
Tecnick
NoiseRegu
−17.65%
−15.32%
−87.43%
−88.43%
CLIC
Clean
−8.59%
−6.26%
−91.49%
−92.92%
CLIC
NoiseRegu
−9.69%
−7.87%
−91.65%
−93.07%
Table 7: Perceptual DMCI-Large: BD-Rate against VTM-22.0. Negative values indicate bitrate savings relative to VTM.
Appendix figures & tables7 assets
Supplementary material from the paper’s appendix.
Appendix
Symbol
Meaning
Inputs and noise
x
clean input image (the distortion target)
η
encoder-side Gaussian noise, η∼N(0,σ2I)
σ
noise scale (the small parameter of the local expansion)
x~=x+η
noisy encoder input
u
uniform proxy for latent quantization, u∼U(−21,21)
Appendix
Table 8: Notation used throughout this appendix. Cross-references to the main text equations are given in the right column where applicable.
δ
QP = 0
9
18
27
36
45
54
63
avg.
0.005
10.28
0 9.33
8.54
10.17
11.99
13.71
14.82
14.70
11.69
0.01
11.14
10.06
9.22
10.42
12.25
13.60
14.68
14.61
12.00
0.02
11.91
10.76
9.71
10.59
12.73
13.32
14.42
14.39
12.23
0.05
12.90
11.49
9.85
10.27
12.42
12.75
14.03
14.22
12.24
Appendix
Table 9: Full sweep: relative reduction (%) in encoder local Lipschitz L^ga on Kodak (DMCI-Base, baseline → noise-regularized). Each cell averages Eq. ( 9 ) ( Φ=ga ) over 24 Kodak images and n=64 random directions. Row averages stay within a 0.6 -pp band as δ varies over an order of magnitude.
δ
QP = 0
9
18
27
36
45
54
63
avg.
0.005
9.33
8.84
8.36
8.75
7.51
7.82
7.44
7.62
8.21
0.01
9.56
9.10
8.46
8.72
7.64
7.97
7.56
7.85
8.36
0.02
9.81
9.30
8.47
8.79
7.83
8.02
7.47
7.86
8.45
0.05
9.99
9.57
8.45
8.56
7.98
8.20
7.38
8.05
8.52
Appendix
Table 10: Full sweep: relative reduction (%) in end-to-end codec local Lipschitz L^f on Kodak (DMCI-Base, baseline → noise-regularized). Each cell averages Eq. ( 9 ) applied to Φ=f over 24 Kodak images and n=64 random directions. Row averages stay within a 0.31 -pp band; the per-QP trend is opposite to that of the encoder, with the strongest reduction at the lowest rate.
Dataset
Variant
bpp
PSNR (dB)
Δ bpp
Δ PSNR (dB)
Kodak
HPCM-Large (baseline)
0.8313
38.022
—
—
HPCM-Large w/ Noise Reg.
0.7965
37.906
−4.19%
−0.116
CLIC
HPCM-Large (baseline)
0.6151
38.691
—
—
HPCM-Large w/ Noise Reg.
0.5867
38.598
−4.61%
−0.093
Tecnick
HPCM-Large (baseline)
0.5613
38.667
—
—
HPCM-Large w/ Noise Reg.
0.5403
38.598
−3.74%
−0.070
Appendix
Table 11: HPCM-Large at a single calibrated operating point on Kodak, CLIC, and Tecnick. We list bpp and PSNR for the reproduced clean-trained baseline and the noise-regularized variant; Δ bpp is the relative bitrate change and Δ PSNR the absolute quality change of the noise-regularized variant relative to the corresponding baseline. We do not aggregate these single-point pairs into a BD-Rate.
Figure 2: RD curves on Kodak . PSNR vs. bitrate (bpp). Curves closer to the upper-left corner are better.
Department of Electronic Engineering, Tsinghua University · Department of Automation, Tsinghua University · State Key Laboratory of Space Network and Communications +1