Teacher-Guided Causal Interventions for Image Denoising: Orthogonal Content-Noise Disentanglement in Vision Transformers

Kuai Jiang, Zhaoyan Ding, Guijuan Zhang, Dianjie Lu, Zhuoran Zheng

Abstract

Conventional image denoising models often inadvertently learn spurious correlations between environmental factors and noise patterns. Moreover, due to high-frequency ambiguity, they struggle to reliably distinguish subtle textures from stochastic noise, resulting in over-removed details or residual noise artifacts. We therefore revisit denoising via causal intervention, arguing that purely correlational fitting entangles intrinsic content with extrinsic noise, which directly degrades robustness under distribution shifts. Motivated by this, we propose the Teacher-Guided Causal Disentanglement Network (TCD-Net), which explicitly decomposes the generative mechanism via structured interventions on feature spaces within a Vision Transformer framework. Specifically, our method integrates three key components: (1) An Environmental Bias Adjustment (EBA) module projects features into a stable, de-centered subspace to suppress global environmental bias (de-confounding). (2) A dual-branch disentanglement head employs an orthogonality constraint to force a strict separation between content and noise representations, preventing information leakage. (3) To resolve structural ambiguity, we leverage Nano Banana Pro, Google's reasoning-guided AI image generation model, to guide a causal prior, effectively pulling content representations back onto the natural-image manifold. Extensive experiments demonstrate that TCD-Net outperforms mainstream methods across multiple benchmarks in both fidelity and efficiency, achieving a real-time speed of 104.2 FPS on a single RTX 5090 GPU.

Introduction: Beyond Correlation in Image Restoration

Most deep learning denoisers are designed as mapping functions that learn to associate noisy inputs with clean outputs through massive data fitting. However, this correlational approach suffers from two critical weaknesses. First, spurious correlations arise when models accidentally link environmental factors—such as specific lighting conditions or sensor-specific biases—with noise patterns. This leads to poor generalization when the model encounters out-of-distribution (OOD) data. Second, high-frequency ambiguity makes it difficult for networks to distinguish between fine, intrinsic textures and stochastic noise, often resulting in over-smoothed images where essential details are lost.

To address these issues, we propose shifting the paradigm from simple pattern matching to causal disentanglement. By explicitly modeling the causal relationship between content, noise, and the environment, we can build more robust and precise restoration systems.

Methodology: The TCD-Net Framework

TCD-Net is built upon a Vision Transformer (ViT) backbone, enhanced by three specific causal interventions designed to separate intrinsic scene content from extrinsic corruption.

Causal diagrams comparing conventional denoising (a) with the proposed TCD-Net (b), highlighting the intervention on environmental bias and the teacher-guided semantic prior.

1. Environmental Bias Adjustment (EBA)

To mitigate the impact of environmental confounders, we introduce the EBA module at the end of each Transformer block. EBA performs a de-centering operation that projects features into a stable subspace, suppressing global environmental biases. This acts as a de-confounding mechanism, ensuring that the subsequent layers process features that are less sensitive to external scene conditions.

2. Dual-Branch Orthogonal Disentanglement

Rather than producing a single output, TCD-Net utilizes a dual-branch architecture that predicts both the restored image and the noise map. We impose a strict orthogonality constraint ($L_{ortho}$) on these two branches. Mathematically, this forces the content and noise representations to be perpendicular in the feature space, creating a "geometric firewall" that prevents texture information from leaking into the noise estimate.

3. Teacher-Guided Causal Prior

To resolve the ambiguity between fine textures and noise, we leverage Google’s Nano Banana Pro (NBP), a reasoning-guided generative model, as a teacher. During the training phase, NBP provides a causal prior that guides the model's content representations toward the natural-image manifold. This semantic guidance helps the network "anchor" its understanding of what constitutes realistic texture versus random noise. Crucially, this teacher is only required during training; the inference model remains a lightweight, single-pass network.

Experimental Results

We evaluated TCD-Net on both synthetic (Gaussian) and real-world noise benchmarks, including CBSD68, Urban100, and the Smartphone Image Denoising Dataset (SIDD).

Efficiency vs. Performance comparison on CBS68 and SIDD datasets. TCD-Net (red star) achieves superior PSNR while maintaining high inference speeds.

Fidelity and Robustness

TCD-Net achieved state-of-the-art PSNR values across multiple datasets. Notably, on texture-heavy benchmarks like Urban100, the model's ability to preserve fine details was significantly superior to traditional methods. On the real-world SIDD benchmark, TCD-Net reached a PSNR of 40.48 dB, outperforming recent high-performance models like NAFNet and MambaIR.

Inference Efficiency

A key advantage of TCD-Net is its computational efficiency. By avoiding the iterative sampling required by diffusion-based models and utilizing an optimized ViT structure, it achieves an inference speed of 104.2 FPS on an NVIDIA RTX 5090. This makes it suitable for real-time applications where both high fidelity and low latency are required.

Conclusion

TCD-Net demonstrates that incorporating causal principles into image restoration can yield significant gains in both robustness and detail preservation. By explicitly disentangling content from noise through orthogonality and leveraging generative priors, we move closer to denoisers that understand the underlying structure of the scenes they restore.

Summary updated

Abstract

Conventional image denoising models often inadvertently learn spurious correlations between environmental factors and noise patterns. Moreover, due to high-frequency ambiguity, they struggle to reliably distinguish subtle textures from stochastic noise, resulting in over-removed details or residual noise artifacts. We therefore revisit denoising via causal intervention, arguing that purely correlational fitting entangles intrinsic content with extrinsic noise, which directly degrades robustness under distribution shifts. Motivated by this, we propose the Teacher-Guided Causal Disentanglement Network (TCD-Net), which explicitly decomposes the generative mechanism via structured interventions on feature spaces within a Vision Transformer framework. Specifically, our method integrates three key components: (1) An Environmental Bias Adjustment (EBA) module projects features into a stable, de-centered subspace to suppress global environmental bias (de-confounding). (2) A dual-branch disentanglement head employs an orthogonality constraint to force a strict separation between content and noise representations, preventing information leakage. (3) To resolve structural ambiguity, we leverage Nano Banana Pro, Google's reasoning-guided AI image generation model, to guide a causal prior, effectively pulling content representations back onto the natural-image manifold. Extensive experiments demonstrate that TCD-Net outperforms mainstream methods across multiple benchmarks in both fidelity and efficiency, achieving a real-time speed of 104.2 FPS on a single RTX 5090 GPU.
BibTeX
@misc{jiang2026teacher,
  author = {Kuai Jiang and Zhaoyan Ding and Guijuan Zhang and Dianjie Lu and Zhuoran Zheng},
  title = {{Teacher-Guided Causal Interventions for Image Denoising: Orthogonal Content-Noise Disentanglement in Vision Transformers}},
  year = {2026},
  eprint = {2603.01140},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2603.01140},
}
Figure