🤖 AI

Perception-based Image Denoising via Generative Compression

This paper proposes a generative compression framework for image denoising that leverages entropy-coded latent representations and perceptual-driven decoders (via conditional WGANs or diffusion models) to achieve realistic texture recovery while providing theoretical guarantees on reconstruction error and decoding probability.

Original authors: Nam Nguyen, Thinh Nguyen, Bella Bose

Published 2026-05-19

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Nam Nguyen, Thinh Nguyen, Bella Bose

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a beautiful, high-resolution photograph of a forest, but someone has thrown a bucket of static-filled snow over it. Your goal is to clean the photo.

The Old Way: The "Blurry Fix"
For a long time, computers tried to fix this by simply averaging out the snow. Think of it like trying to smooth out a bumpy road by running a giant, heavy roller over it. The bumps (noise) are gone, but so are the interesting details like the texture of the tree bark or the individual leaves. The result is a clean but "over-smoothed" picture that looks like a plastic painting rather than a real photo. This is what the paper calls "distortion-driven" methods.

The New Idea: The "Smart Re-imagining"
This paper proposes a different approach called Perception-Based Denoising. Instead of just smoothing things out, the computer tries to "re-imagine" what the clean photo should look like based on what it knows about real-world images.

The authors use a clever trick called Generative Compression. Here is the analogy:

The Compression (The Summary): Imagine you have a messy, noisy room. Instead of trying to clean every single speck of dust, you take a quick, compressed mental snapshot of the room's structure (where the bed is, where the window is). You throw away the messy details (the noise) and keep only the essential "blueprint." In the paper, this is done by an Encoder that turns the noisy image into a compact, efficient code (a latent representation).
The Reconstruction (The Artist): Now, you give this blueprint to a highly skilled artist (the Decoder). The artist doesn't just copy the blueprint; they use their knowledge of what real rooms look like to fill in the missing details. They know that windows usually have glass and beds have sheets. They "hallucinate" (in a good way) the fine textures that the noise destroyed, making the image look realistic again.

Two Different Artists
The paper introduces two specific "artists" (methods) to do this job:

The WGAN Artist (The Adversarial Critic): This method uses a "Game of Two Players." One player (the Generator) tries to create a clean image from the blueprint. The other player (the Discriminator) acts like a strict art critic, trying to spot if the image is fake or real. They play this game over and over until the Generator becomes so good that the critic can't tell the difference. This ensures the final image looks incredibly realistic, not just mathematically correct.
The Diffusion Artist (The Step-by-Step Sculptor): This method is like a sculptor starting with a block of noisy clay and slowly chipping away the noise, step-by-step, guided by the blueprint. It starts with pure chaos and gradually refines it into a clear image, ensuring that every tiny detail fits perfectly with the overall structure.

The Safety Net (The Math)
The authors didn't just build these tools; they also wrote a rulebook (mathematical proofs) to guarantee that their method works. They proved that if the noise isn't too crazy, their "blueprint and artist" system will always produce a result that is close to the original, and they calculated exactly how likely it is to make a mistake.

The Results
When they tested this on real photos (like nature scenes) and even tricky scientific photos (like microscope images of cells), the results were impressive:

Old methods made the images look smooth and blurry.
Their methods kept the images sharp and full of texture, looking much more like what a human eye expects to see.

In short, instead of just "cleaning" the noise, this paper teaches the computer to reconstruct the image, filling in the missing pieces with realistic details, resulting in photos that look natural and vibrant again.

Technical Summary: Perception-Based Image Denoising via Generative Compression

1. Problem Statement

Image denoising aims to remove noise while preserving structural details and perceptual realism. However, traditional methods optimized for pixel-wise distortion metrics (e.g., Mean Squared Error or PSNR) often produce over-smoothed reconstructions, particularly under strong noise or distribution shifts. This occurs because minimizing distortion alone is insufficient for perceptual quality; the Rate-Distortion-Perception (RDP) trade-off formalizes that minimizing distortion without considering perceptual constraints leads to a loss of fine textures and natural image statistics.

Existing perceptual restoration methods, such as Conditional Wasserstein GANs (WGANs), can suffer from unstable training and inadequate capture of natural image diversity. Conversely, diffusion models offer stable training and strong mode coverage but require integration with compression principles to effectively handle the structured nature of noise versus natural signals.

2. Methodology

The paper proposes a generative compression framework where denoising is achieved by reconstructing images from entropy-coded latent representations. The core premise is that natural signals exhibit low complexity, whereas noise is unstructured. By operating a compressor at a distortion level matched to the noise strength, high-complexity noise is suppressed while structured content is preserved.

The authors introduce two complementary instantiations of this framework:

2.1. Conditional WGAN-Based Compression Denoiser (CGanDeCompress)

This approach integrates learned lossy compression with conditional adversarial training.

Architecture: An encoder ( $E$ ) maps a noisy input $y$ to a quantized latent representation $z$ . A decoder ( $G$ ) reconstructs the denoised image $\hat{x}$ from $z$ .
Rate Control: The latent $z$ is entropy-coded using a learned probability model $P$ , yielding a rate $r(z) = -\log P(z)$ .
Training Objective: The system minimizes a combined loss function:
$L_{EGP} = \mathbb{E}_y[\lambda r(z) + d(x, \hat{x}) - \beta \log D(\hat{x}, z)]$
where $d(x, \hat{x})$ combines MSE and Learned Perceptual Image Patch Similarity (LPIPS) loss. A conditional WGAN discriminator $D$ distinguishes reconstructed images from clean images conditioned on $z$ , promoting perceptual realism.
Control: Parameters $\lambda$ and $\beta$ explicitly control the rate-distortion-perception (RDP) trade-off.

2.2. Conditional Diffusion-Based Reconstruction (DiffDeCompress)

This approach leverages diffusion models as generative priors guided by compressed latents.

Architecture: A stochastic encoder maps $y$ to a compact latent $z$ . A decoder models the conditional distribution $p(x_0|z)$ via a reverse diffusion process.
Process: The reverse process is conditioned on $z$ , guiding the denoising trajectory toward semantically consistent reconstructions while intermediate latent variables allow for the synthesis of fine-scale details.
Training Objective: The model optimizes a variational upper bound on the diffusion negative log-likelihood, augmented with an explicit LPIPS loss:
$L = \rho L_p + (1-\rho)L_c$
where $L_p$ is the LPIPS loss and $L_c$ includes the rate constraint and the diffusion reconstruction loss.

3. Theoretical Contributions

The paper establishes non-asymptotic guarantees for compression-based maximum-likelihood (ML) denoising under additive white Gaussian noise (AWGN).

Reconstruction Error Bounds: The authors derive upper bounds on the reconstruction error $\|x - \hat{x}\|$ in terms of the compression rate $R$ , distortion $D(P)$ , and perception level $P$ . The bound indicates that with high probability, the error is close to the theoretical distortion-perception limit $D(P)$ .
Decoding Error Probability: The paper provides a bound on the probability of the ML denoiser selecting an incorrect codeword, showing that this probability decreases exponentially with the compression rate $R$ .
Connection to RDP Theory: The theoretical framework explicitly connects denoising performance to the RDP theory, demonstrating that the proposed methods operate within the theoretical limits of distortion-perception trade-offs.

4. Experimental Results

The methods were evaluated on synthetic Gaussian noise (KODAK and DIV2K datasets) and real-world noise (Fluorescence Microscopy Dataset and Smartphone Image Denoising Dataset - SIDD).

Synthetic Noise:
- Perceptual Quality: Both proposed methods consistently outperformed classical (BM3D) and learning-based (N2C, N2N, Deep Decoder) baselines in perceptual metrics (FID, LPIPS, PI).
- Distortion Trade-off: While supervised methods like N2C achieved higher PSNR/SSIM, they exhibited significantly higher FID and LPIPS scores, indicating over-smoothed outputs. The proposed methods maintained competitive distortion performance while achieving superior perceptual realism.
- Specifics: At $\sigma=15$ , CGanDeCompress achieved the lowest FID and LPIPS. At higher noise levels ( $\sigma=50$ ), DiffDeCompress achieved the best perceptual quality (lowest PI) on KODAK, while CGanDeCompress excelled in distributional similarity (FID) on DIV2K.
Real-World Noise:
- On the FMD and SIDD datasets, CGanDeCompress demonstrated robustness to domain shifts and complex real-world noise.
- FMD: LPIPS improved from 0.439 (noisy) to 0.035, with PSNR increasing from 28.45 dB to 36.71 dB.
- SIDD: LPIPS reduced from 0.646 to 0.093, and PSNR increased from 26.35 dB to 32.70 dB, with significant gains in structural similarity (SSIM).

5. Significance and Claims

The paper claims that the proposed generative compression framework successfully bridges the gap between distortion-driven and perception-driven denoising. By interpreting denoising as a projection onto a low-complexity signal class induced by a lossy codec, the method achieves:

Explicit Control: The ability to explicitly manage the RDP trade-off through learned compression models.
Perceptual Fidelity: The generation of realistic textures and fine details that are often lost in MSE-optimized methods.
Theoretical Grounding: The establishment of non-asymptotic bounds linking compression rate to denoising error, validating the approach through information-theoretic principles.
Robustness: Consistent performance across synthetic benchmarks and challenging real-world scenarios involving domain shifts and complex noise distributions.

The authors conclude that generative compression offers a robust paradigm for perception-based image denoising, achieving improved perceptual fidelity without sacrificing competitive distortion performance.