Continuous Degradation Modeling via Latent Flow Matching for Real-World Super-Resolution

Hyeonjae Kim, Dongjin Kim, Eugene Jin, Tae Hyun Kim

TL;DR

DegFlow is a new framework that models real-world image degradation (blur, noise, artifacts) as a continuous trajectory in a latent space using flow matching. By learning from a small set of discrete examples, it can synthesize highly realistic low-resolution training data at any arbitrary scale, significantly boosting the performance of super-resolution models on actual photographs.

What Problem Does This Paper Address?

Most deep learning models for Super-Resolution (SR)—the task of upscaling images—are trained on "clean" synthetic data, typically generated by simple bicubic downsampling. However, real-world images captured by DSLRs or smartphones suffer from complex, non-linear degradations like sensor noise, lens blur, and compression artifacts.

When models trained on clean data meet messy real-world images, they fail. To fix this, researchers usually have to painstakingly collect "paired" datasets (taking the same photo at different zoom levels), which is labor-intensive and limited to fixed scales (e.g., exactly 2x or 4x). There is a desperate need for a way to generate "authentic-looking" low-resolution (LR) images from high-resolution (HR) sources at any scale to provide better training data.

Methods

The authors introduce DegFlow, a two-stage pipeline inspired by latent diffusion and flow matching:

  1. Residual Autoencoder (RAE): First, they train an autoencoder to compress images into a compact latent space. A clever trick here is the use of multi-scale skip connections that pass high-frequency details from the HR image directly to the decoder. This forces the "latent code" to focus strictly on representing the degradation (the noise and blur) rather than the underlying image content.
  2. Latent Flow Matching (LFM): Instead of just interpolating between a sharp image and a blurry one, DegFlow treats degradation as a continuous path (a "flow") in the latent space.
  3. Natural Cubic Splines: Unlike previous methods that used simple linear interpolation, DegFlow uses cubic splines to model the trajectory. This accounts for the non-linear nature of how images lose quality as they are downscaled.
  4. 3rd-order Taylor Approximation: To train the model on scales where no ground truth exists (like 3.2x), they use a Taylor expansion to "project" the predicted latent toward the nearest known scale (like 4x) for perceptual supervision.

At inference time, you give DegFlow a single HR image and a target scale, and it "flows" the latent representation to that specific degradation level, producing a realistic LR counterpart.

Datasets

The authors utilized several key benchmarks for training and evaluation:

  • RealSR (V2): A dataset containing paired HR-LR images from Canon and Nikon cameras at x2, x3, and x4 scales.
  • RealArbiSR: Used to test the model's ability to handle arbitrary, non-integer scaling factors.
  • DIV2K: A high-quality dataset used as a source for "external" HR images to show that DegFlow can generate useful training data even from images it hasn't seen before.

Key Insights and Ideas

  • Degradation is a Trajectory: The core "aha" moment is treating the transition from sharp to blurry not as a jump, but as a smooth, continuous movement through a mathematical space.
  • Isolating the "Mess": By using skip connections for the "clean" features, the latent space becomes a dedicated "degradation manifold." This makes the flow matching task much easier because the model doesn't have to learn how to reconstruct a house or a tree—it only learns how to apply the "look" of a Nikon sensor at 3x zoom.
  • Smoothness Matters: The shift from linear interpolation to cubic splines is vital. Real-world physics (optics, electronics) rarely moves in straight lines; the spline ensures the first and second derivatives of the degradation path are continuous, leading to more stable and realistic results.

Key Results

DegFlow consistently outperformed existing state-of-the-art (SOTA) methods like InterFlow and Real-ESRGAN:

  • Better Fidelity: On the RealSR x3 test set, SR models trained with DegFlow data achieved higher PSNR (signal-to-noise ratio) and SSIM (structural similarity) than those trained on other synthetic datasets.
  • Arbitrary Scaling: For models like MetaSR and LIIF (which can upscale to any size), DegFlow provided much better training supervision. For example, on the RealArbiSR benchmark at x3.5 scale, DegFlow improved PSNR by nearly 0.5 dB over the previous best method.
  • Visual Realism: Qualitative results show that DegFlow-generated images capture the specific "texture" of real camera noise and blur, whereas other methods often produce "plastic-looking" or overly sharpened artifacts.

Limitations and Future Work

The authors note a primary limitation: while DegFlow synthesizes the appearance of real-world degradation, it currently does so at the same spatial resolution as the input HR image. To train an actual SR model, they still have to apply a final bicubic downsampling step to physically resize the image.

Future work could involve integrating the physical downsampling directly into the flow process, potentially creating an even more seamless end-to-end degradation model that doesn't rely on any classical resizing operators.

Summary updated

Abstract

While deep learning-based super-resolution (SR) methods have shown impressive outcomes with synthetic degradation scenarios such as bicubic downsampling, they frequently struggle to perform well on real-world images that feature complex, nonlinear degradations like noise, blur, and compression artifacts. Recent efforts to address this issue have involved the painstaking compilation of real low-resolution (LR) and high-resolution (HR) image pairs, usually limited to several specific downscaling factors. To address these challenges, our work introduces a novel framework capable of synthesizing authentic LR images from a single HR image by leveraging the latent degradation space with flow matching. Our approach generates LR images with realistic artifacts at unseen degradation levels, which facilitates the creation of large-scale, real-world SR training datasets. Comprehensive quantitative and qualitative assessments verify that our synthetic LR images accurately replicate real-world degradations. Furthermore, both traditional and arbitrary-scale SR models trained using our datasets consistently yield much better HR outcomes.
BibTeX
@misc{kim2026continuous,
  author = {Hyeonjae Kim and Dongjin Kim and Eugene Jin and Tae Hyun Kim},
  title = {{Continuous Degradation Modeling via Latent Flow Matching for Real-World Super-Resolution}},
  year = {2026},
  eprint = {2602.04193},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2602.04193},
}
Figure