TT-net: Quantum Inspired Tensor Network Denoising in Conditional GANs

Abstract: Developed as a workhorse for classical simulations of quantum algorithms and quantum many-body systems, Tensor Network methods have entered the scientific mainstream in quantum physics. Among various types of tensor networks, Tensor Trains (commonly know as Matrix Product States in the quantum computing community) have already found applications in machine learning. These methods often rely on a powerful linear algebra tool called the Singular Value Decomposition (SVD). Several conditional GAN architectures for image denoising incorporate SVD as a single-cut decomposition step applied to generator feature maps. In this work we introduce TT-Net, which replaces the per-channel SVD denoising block with a two-cut tensor-train decomposition capable of accessing cross-channel information directly, a capability absent from contemporary alternatives. In a controlled comparison differing only in this decomposition mechanism, TT-Net outperforms SVD-Net on PSNR and SSIM across all three noise types tested (Gaussian, motion blur, and salt-and-pepper), supporting the hypothesis that cross-channel access improves denoising quality. Training-dynamics analysis further shows that TT-Net's adversarial loss term consistently saturates to a stagnant state across all three noise types, more so than SVD-Net's, while reconstruction quality continues to improve regardless, raising an open question about the adversarial component's contribution that this work identifies but does not resolve. Furthermore, for Gaussian noise our method outperforms both the EigenGAN and the state of the art Pix2pix method which does not assume any linear algebra decompositions and does not retain any linear algebra information. Our manuscript shows how quantum inspired tools can be used as practical real world feature filters for deep learning applications.

• The paper introduces TT-Net, a tensor-train denoising technique in conditional GANs, achieving superior performance in image denoising by enabling cross-channel access that SVD-based methods lack
• Across various noise types, TT-Net shows improvements of +2.68 dB / +0.0513 SSIM for Gaussian, +8.32 dB / +0.3798 SSIM for motion blur, and +0.99 dB / +0.0165 SSIM for salt-and-pepper noise, when compared to previous approaches
• Training dynamics analysis questions traditional adversarial contributions to performance and suggests future experiments could investigate the necessity of the GAN framework because reconstruction quality improvements persist after adversarial feedback saturates
• follow_up_questions
• What architectures can TT-Net be combined with to enhance image reconstruction?
• How does varying the rank in tensor-train decomposition affect noise reduction performance?
• How does TT-Net scale up through exploratory comparisons to higher-resolution images?
• Are there concurrent textures or textures involved beyond the video GANs that can innovate dynamically in future implementations?
• Find recent papers about adversarial contributions to denoising performance.
• find
• Using tensor network techniques we can contourally compile video GANs to enhance separate diffuse high-definition imagery
• Main hypothesis of TT-Net: since channels are isolated, there might be independence with low-rank channel transitions between channels leading to independence vis a vis the transfer steerability with camera,
• Overall reconstruction quality visually improves for TT-Net.

Motivation and contribution

TT-Net addresses a specific architectural gap in conditional GAN-based image denoising. Prior architectures embed a singular value decomposition (SVD) inside the generator—EigenGAN adjusts generated feature-map singular values toward target-image values via a supervised loss term, while SVD-Net uses SVD purely as a per-channel energy filter after each encoder downsampling stage. In both cases the SVD is a single-cut decomposition: the three-axis feature map (channels × height × width) is unfolded into one matrix once, so whichever axes are merged can never be examined separately. SVD-Net's per-channel scheme is therefore structurally incapable of exploiting inter-channel correlations, at any threshold.

TT-Net replaces this block with a two-cut tensor-train decomposition—the machine-learning analogue of a Matrix Product State—applied at the same three insertion points as SVD-Net's filter, under otherwise identical training conditions. The stated contributions are: (1) a tensor-train denoising block with direct cross-channel access; (2) an adaptive per-cut thresholding scheme matching SVD-Net's retained-energy target for fairness; and (3) a four-model comparison across Gaussian noise, motion blur, and salt-and-pepper corruption on CIFAR-10 resized to 64×64.

Method

For a feature map $\mathcal{X} \in \mathbb{R}^{C \times H \times W}$, the first cut reshapes into $M_1 \in \mathbb{R}^{C \times HW}$ so that all channels of an image appear as rows of a single matrix and are compared against each other during decomposition—an operation impossible under per-channel SVD. An adaptive rank $\chi_1$ retains dominant directions, which correspond to patterns of agreement across channels; residual channel-specific variation is discarded. The remainder $R_1$ is reshaped to merge channel and height, and a second SVD over $M_2 \in \mathbb{R}^{(\chi_1 H) \times W}$ captures horizontal spatial regularities shared jointly across retained channel directions and height positions. The filtered map is reconstructed by contracting three cores:

$$\widetilde{\mathcal{X}}[i,j,k] = \sum_{a=1}^{\chi_1}\sum_{b=1}^{\chi_2}\text{Core}_1[i,a]\,\text{Core}_2[a,j,b]\,\text{Core}_3[b,k].$$

Because the second cut operates on the first cut's remainder, per-cut thresholds compound multiplicatively; setting both to 0.9 would retain only ~0.81 of the energy overall. The authors therefore set $\theta_1 = \theta_2 = \sqrt{0.9} \approx 0.9487$, matching SVD-Net's nominal $\theta = 0.9$. Since ranks are chosen as the smallest sufficient discrete value, actual retained energy typically exceeds this target—an acknowledged asymmetry, though a mild one.

The design isolates the decomposition mechanism as the sole variable between TT-Net and SVD-Net: identical generator/discriminator structure, non-saturating adversarial loss, and reconstruction loss blending MSE and MS-SSIM ($\delta = 0.84$). TT-Net adopts SVD-Net's optimizer settings unchanged rather than introducing independently tuned hyperparameters. Two disclosed deviations matter for interpretation: an extra encoder downscaling stage was added to fit GPU constraints (inherited by both models), making their encoder shallower than Pix2pix/EigenGAN's four-stage bottleneck; consequently, Pix2pix and EigenGAN serve only as contextual baselines, not controlled comparisons.

Quantitative results

Within the controlled comparison, TT-Net outperforms SVD-Net on every metric and every noise type:

Model	Gauss. PSNR	Gauss. SSIM	Blur PSNR	Blur SSIM	S&P PSNR	S&P SSIM
Pix2pix	25.52	0.7466	36.18	0.9643	41.79	0.9881
EigenGAN	25.18	0.7545	26.71	0.8398	34.24	0.9513
SVD-Net	23.60	0.7455	18.02†	0.4260†	28.98	0.8761
TT-Net	26.28	0.7968	26.34	0.8058	29.97	0.8926

†SVD-Net's motion-blur result reflects late-training discriminator collapse, not convergence.

Gains over SVD-Net are +2.68 dB / +0.0513 SSIM (Gaussian), +8.32 dB / +0.3798 SSIM (motion blur), and +0.99 dB / +0.0165 SSIM (salt-and-pepper). The authors correctly flag that the motion-blur gap is partly confounded by SVD-Net's training failure; the Gaussian and salt-and-pepper gaps, where no instability occurred in either run, are the cleaner evidence for the cross-channel hypothesis.

The strongest headline claim concerns Gaussian noise: TT-Net achieves the best PSNR and SSIM of all four models, exceeding Pix2pix by 0.76 dB PSNR and 0.0502 SSIM despite Pix2pix's deeper encoder. Because SVD-Net underperforms both baselines here, this lead cannot be attributed to encoder depth and more directly reflects the decomposition mechanism. Qualitatively, TT-Net preserves natural coloring where SVD-Net produces green-tinted saturation on Gaussian inputs, and remains coherent on motion blur where SVD-Net outputs corrupted patches.

Training dynamics and the role of the adversary

A notable secondary finding emerges from training-dynamics analysis. In every run except SVD-Net's motion-blur failure, discriminator outputs $D(\text{real})$ and $M_1 \in \mathbb{R}^{C \times HW}$0 converge rapidly toward ~0.500—a state offering no informative gradient—and TT-Net reaches this saturation faster and more consistently than SVD-Net (within two epochs on Gaussian). Critically, PSNR and SSIM continue improving throughout training even after adversarial feedback saturates; TT-Net's total generator loss keeps declining, driven almost entirely by the still-improving MS-SSIM/MSE reconstruction terms.

This raises a pointed question about the adversarial component's actual contribution to denoising performance—one the paper identifies explicitly but does not resolve. Confirming that reconstruction-quality gains persist without any adversarial term would require ablation runs with the adversarial loss removed entirely, which the authors defer to future work. Until that experiment exists, TT-Net's framing as a conditional GAN rests on an unverified premise about whether its discriminator contributes anything beyond initialization effects.

Limitations and open questions

Three limitations bound the conclusions, all acknowledged by the authors. First, the motion-blur comparison is confounded by SVD-Net's discriminator-driven collapse (D(real) climbing monotonically to 0.876 from epoch 7 onward while D(fake) falls to 0.106), so the largest reported gap cannot be cleanly attributed to the mechanism. Second, test metrics use final-epoch checkpoints rather than best-validation checkpoints, which may slightly understate or misorder converged quality. Third, evaluation is confined to a single dataset (CIFAR-10) with synthetic noise; generalization to real-world noise or higher resolutions is untested. Open questions include whether the adversarial loss is necessary at all, how per-block retained rank correlates with noise difficulty, and whether the mechanism transfers beyond CIFAR-10-scale images.

Conclusion

TT-Net demonstrates that upgrading a single-cut SVD activation filter to a two-cut tensor-train decomposition yields consistent, measurable improvements in conditional-GAN image denoising, supporting the hypothesis that cross-channel access is the operative factor—with the caveat that the cleanest evidence comes from the Gaussian and salt-and-pepper conditions. Its best-overall performance on Gaussian noise, ahead of architecturally deeper baselines, is the paper's most consequential empirical claim. Equally significant is the training-dynamics observation that adversarial losses saturate while reconstruction quality improves regardless, leaving unresolved whether the GAN framework itself is load-bearing in this architecture.