Unpaired Image Restoration for Real World Degradation

This direction is merged into another direction: paired data dependency (ground-truth data scarcity). Its last version stays readable and citable here.

Version 1 of 125 papers25 conceptsfrozen 2026-10-04

In an ideal world, training a model to fix a blurry or foggy image would be simple: you would show it millions of "before and after" pairs. By comparing a degraded image directly to its pristine ground truth, a neural network can learn exactly which pixels to change.

In the real world, these pairs almost never exist. You cannot take a photo of a mountain in thick fog and then instantly "turn off" the weather to capture the exact same scene in clear air. You cannot take a high-resolution photo of a person from a 1920s street scene to train a restoration model for historical archives. This "paired data bottleneck" has led researchers to develop a suite of techniques for unpaired image restoration, where models learn to clean images by looking at two independent sets—one of "dirty" images and one of "clean" images—without any direct correspondence between them.

Closing the Loop with Dual Regression

If we don't have a ground truth image to compare against, how do we know if a restoration is accurate? One of the most powerful intuitions in this field is the "closed-loop" constraint.

In Dual Regression Networks (DRN), Jian Chen and colleagues argue that a restoration mapping should be reversible. If a model upscales a low-resolution image to high-resolution, the resulting image, when downsampled, should look exactly like the original input. This creates a self-supervised signal: the input itself becomes the ground truth for the "round trip."

Similarly, Guosheng Yin applied this to dehazing in Efficient Unpaired Image Dehazing by using cyclic supervision. By ensuring that a restored image can be "re-hazed" back to its original state, the model is forced to preserve the underlying geometry and structure of the scene rather than just hallucinating a pretty picture.

A more recent shift moves away from simple "round trips" toward modeling the underlying distribution of what "clean" images look like. If a model understands the statistical structure of sharp, clear photos, it can treat restoration as a process of "walking" a degraded image toward that clean distribution.

This is the core intuition behind score-based and diffusion models. In Real-world defocus deblurring, Yuhao Li demonstrates that by training only on the distribution of sharp images, a model can iteratively refine a blurry input until it fits the "sharp" manifold.

To make this process faster and more physically grounded, Jingjun Yi introduced the Degradation-aware Dynamic Schrödinger Bridge (DDSB). Instead of a slow diffusion process, the Schrödinger Bridge finds the most efficient probabilistic path between the "degraded" distribution and the "clean" distribution. By incorporating a learned physical model of the degradation (like how light scatters in fog), the model ensures the restoration path is not just statistically likely, but physically plausible.

Disentangling Content from Degradation

Another way to solve the unpaired problem is to treat an image as a combination of two separate parts: the content (what is in the scene) and the degradation (the blur, rain, or haze).

Yukun Huang explored this in Real-World Person Re-Identification by using "disentangled representation learning." The goal is to strip away the "noise" of low resolution while keeping the "identity" of the person intact.

Contrastive learning has become a favorite tool for this disentanglement. In CCLGAN, Chen Zhao uses contrastive losses to ensure that a derained image maintains the same semantic meaning as the rainy input. By forcing the model to recognize that a "rainy tree" and a "clean tree" share the same essence, the model learns to remove only the rain streaks without altering the tree itself.

Learning in Latent Space

Sometimes, the best way to bridge the gap between unpaired images is to stop looking at pixels entirely. Longguang Wang proposed DictSR, which learns "coupled dictionaries." Instead of mapping a low-res pixel grid to a high-res one, the model learns a shared latent space where a low-res patch and its high-res counterpart share the same "code."

This approach is particularly useful for tasks like modernizing old photos, where Agus Gunawan showed that we can "borrow" styles from modern reference images. By mapping historical textures and modern aesthetics into a shared space, the model can translate an old photo into a modern one as if it were translating between two languages.

Why It Works: The Power of Priors

The common thread across these 24 papers is the use of priors—assumptions about the world that substitute for missing ground truth data. Whether it is the geometric prior of depth Yin et al., the physical prior of the Retinex theory Dehaze-RetinexGAN, or the frequency-domain prior that haze lives in the amplitude spectrum Chengxu Liu, these models succeed because they don't just guess. They use the structure of the physical world to constrain the infinite possibilities of what a "clean" image could be.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_unpaired_image_restoration_for_real_worl_d2c2a902,
  title = {Unpaired Image Restoration for Real World Degradation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/unpaired-image-restoration-for-real-world-degradation/rdr_d2c2a90293f94025b403ea9701a8c6a0},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_unpaired_image_restoration_for_real_worl_d2c2a902_v1,
  title = {Unpaired Image Restoration for Real World Degradation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/unpaired-image-restoration-for-real-world-degradation/rdr_d2c2a90293f94025b403ea9701a8c6a0/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}