Rectified Flow Degradation Module in SR

Updated 8 July 2026

RFDM is a degradation modeling module that synthesizes realistic low-resolution images by learning a continuous, invertible flow between degradation-transformed and real images.
It employs a UNet-parameterized velocity field to simulate degradation trajectories, enabling the pairing of synthetic LR images with high-resolution counterparts for supervised training.
Empirical evaluations show that combining RFDM with Fourier-domain guidance enhances super-resolution performance metrics such as PSNR, SSIM, LPIPS, and FID.

Rectified Flow Degradation Module (RFDM) most specifically denotes the degradation-modeling component introduced for unsupervised real-world super-resolution in "Unsupervised Real-World Super-Resolution via Rectified Flow Degradation Modelling" (Zhou et al., 10 Aug 2025). In that setting, RFDM is a rectified-flow-based module that introduces degradation-transformed low-resolution (DT-LR) images as intermediaries and models the degradation trajectory from DT-LR to real low-resolution (LR) images in a continuous and invertible manner. The resulting synthetic LR images are paired with given high-resolution (HR) images to train off-the-shelf super-resolution networks, with the stated aim of reducing the domain gap between synthetic and real-world degradations (Zhou et al., 10 Aug 2025).

1. Terminological scope and acronym disambiguation

The acronym RFDM is not unique across recent rectified-flow literature. In real-world super-resolution, it refers to the Rectified Flow Degradation Module proper. In other contexts, closely related or entirely different expansions are used, and conflating them can obscure the role of degradation modeling versus restoration or video editing (Zhou et al., 10 Aug 2025).

Usage	Paper	Meaning
RFDM	"Unsupervised Real-World Super-Resolution via Rectified Flow Degradation Modelling" (Zhou et al., 10 Aug 2025)	Rectified Flow Degradation Module
RFDM-equivalent	"WaterFlow: Explicit Physics-Prior Rectified Flow for Underwater Saliency Mask Generation" (Li et al., 14 Oct 2025)	The Underwater Physical Prior Module is described as the functional equivalent of a Rectified Flow Degradation Module
RFDM	"RFDM: Residual Flow Diffusion Model for Efficient Causal Video Editing" (Salehi et al., 6 Feb 2026)	Residual Flow Diffusion Model

In the super-resolution paper, RFDM is a degradation model used to synthesize realistic LR observations from unpaired data. In "WaterFlow," the module with analogous functional status is the Underwater Physical Prior Module (UPPM), which encodes degradation cues such as attenuation, color shift, and backscatter as structured information rather than interference (Li et al., 14 Oct 2025). By contrast, in the causal video editing paper, RFDM names a residual forward-process formulation for frame-to-frame editing and does not denote a degradation synthesis module (Salehi et al., 6 Feb 2026).

A common misconception is therefore that every occurrence of "RFDM" in rectified-flow work refers to the same mechanism. The available papers indicate otherwise: the same acronym spans degradation modeling, physics-prior injection, and residual diffusion formulations.

2. Problem formulation in unsupervised real-world super-resolution

Within unsupervised real-world super-resolution, RFDM is motivated by a specific training mismatch: real LR images exhibit complex and unknown degradations, while conventional supervised SR training typically relies on synthetic LR-HR pairs whose degradation process is substantially simpler. The cited paper characterizes this as a significant domain gap and positions RFDM as a method for effectively capturing and modeling real-world degradation from unpaired LR and HR image sets (Zhou et al., 10 Aug 2025).

The module is built on the observation that repeated up-down sampling of LR images leads to similar degradation distributions. This observation underlies the introduction of DT-LR images, which serve as an intermediate, more tractable source domain. Real LR images remain the target domain. A plausible implication is that the module does not attempt to hand-design the degradation process; instead, it learns a transport trajectory between two empirically constructed distributions that bracket the desired realism.

This formulation differs from restoration-oriented rectified-flow methods. "IR-Flow: Bridging Discriminative and Generative Image Restoration via Rectified Flow" directly establishes a linear transport flow between degraded and clean image domains for restoration (Fan et al., 21 Apr 2026), whereas RFDM in super-resolution models the forward degradation needed to synthesize training pairs (Zhou et al., 10 Aug 2025). The distinction is operationally important: the former restores images, while the latter manufactures realistic degradations so that a downstream SR network can be trained in supervised fashion.

3. Architectural organization and synthesis pipeline

RFDM operates in a staged pipeline. Real LR images are first subjected to repeated up-down sampling, exemplified in the paper by 10 bilinear rescale cycles, to produce DT-LR images. RFDM then models a flow between DT-LR as source and real-LR as target. The flow is parameterized via a UNet architecture that models the time-dependent velocity field (Zhou et al., 10 Aug 2025).

The full synthesis pathway in the paper also includes the Fourier Prior Guided Degradation Module (FGDM). FGDM leverages structural information embedded in Fourier phase components to ensure more precise modeling of real-world degradation. In synthesis mode, HR images are downsampled by bilinear interpolation to create initial LR images; these are first processed by FGDM and then passed through RFDM, producing final synthetic LR images with real-world degradation. The synthetic LR images are paired with the given HR images to train off-the-shelf SR networks (Zhou et al., 10 Aug 2025).

This division of labor assigns complementary functions to the two modules. RFDM is responsible for the learned degradation flow, while FGDM contributes Fourier-domain structure guidance. The ablation summary in the paper states that RFDM only yields lower performance due to lack of structural information in DT-LR inputs; FGDM only improves performance due to Fourier domain structure guidance; and the combined configuration gives the best performance, indicating complementarity (Zhou et al., 10 Aug 2025).

The architectural idea has a broader analogue in "FlowIE: Efficient Image Enhancement via Rectified Flow," where degradation modules are described as conditions for learning a straight path from an elementary distribution to high-quality images (Zhu et al., 2024). This suggests a broader rectified-flow design pattern in which degradation modeling is not necessarily a preprocessing artifact but a condition or transport target integrated into the learned trajectory itself.

4. Mathematical formulation and rectified-flow mechanics

RFDM formalizes degradation modeling as an ordinary differential equation

$dZ_t = v(Z_t,t)\,dt,$

where $Z_0$ is drawn from the DT-LR image distribution and $Z_1$ from the real-LR image distribution. The time-dependent velocity field $v(\cdot,t)$ guides trajectories from the source distribution to the target distribution (Zhou et al., 10 Aug 2025).

The module adopts a rectified-flow formulation in which

$v(Z_t,t)=\mathbb{E}[Z_1-Z_0\mid Z_t],$

with linear interpolation

$Z_t=tZ_1+(1-t)Z_0,\qquad t\in[0,1].$

Training minimizes a flow-matching mean-squared objective,

$\min_\theta \int_0^1 \mathbb{E}\left[\|(Z_1-Z_0)-v_\theta(Z_t,t)\|^2\right]dt.$

The paper further injects Gaussian noise into the DT-LR input,

$X_0:=\bar{X}+\lambda n,\qquad n\sim\mathcal{N}(0,I),$

to avoid singularities and ensure good coverage in high/low-dimensional corners of image space (Zhou et al., 10 Aug 2025).

Sampling is performed by discrete ODE integration of the learned velocity field. After $K$ steps, the model produces a degradation-transformed image matching the real LR distribution. The ablation summary reports that increasing $K$ improves performance until a plateau and identifies approximately 20 steps as optimal (Zhou et al., 10 Aug 2025).

The paper characterizes the learned transformation as continuous and, under ideal integration, invertible. In context, this is intended to distinguish RFDM from one-shot mappings and from adversarial or diffusion approaches that either introduce artifacts, suffer instability, or do not explicitly model a continuous degradation trajectory (Zhou et al., 10 Aug 2025). Relative to generic rectified-flow expositions, this use is noteworthy because the terminal state is not a clean image or segmentation target but a realistic degraded observation.

5. Empirical behavior, ablations, and downstream effect

The empirical role of RFDM is evaluated through downstream SR performance on real-world datasets. The paper states that SR networks including SwinIR, Real-ESRGAN, and StableSR, when trained on data synthesized using RFDM + FGDM, consistently outperform baselines including Real-ESRGAN, UDDM, and Syn-Real on RealSR and DRealSR across PSNR, SSIM, LPIPS, and FID. One explicit example is given for SwinIR on RealSR: UDDM attains 26.732 / 0.7913, whereas the proposed method attains 27.022 / 0.7981 (Zhou et al., 10 Aug 2025).

The ablations isolate several factors. First, module composition matters: RFDM only is inferior, FGDM only improves performance, and using both is best. Second, noise level matters: no noise ( $Z_0$ 0) yields minimal RFDM effect, moderate noise ( $Z_0$ 1) yields the best PSNR/SSIM, and excessive noise degrades visual quality. Third, solver length matters: increasing the number of ODE steps improves performance until a plateau, with approximately 20 steps reported as optimal. Fourth, DT-LR construction matters: bilinear up-down sampling for 10 cycles provides the best trade-off between information retention and degradation similarity, while Bicubic and Lanczos are reported as less effective (Zhou et al., 10 Aug 2025).

The qualitative evidence follows the same pattern. Synthetic LR images generated via RFDM + FGDM are described as much closer visually to true real-LR images than alternatives, and SR results trained on such synthetic pairs are described as having fewer artifacts, better structure/detail, and less blurring than baseline methods (Zhou et al., 10 Aug 2025). Taken together, these results position RFDM less as an end-user restoration module than as a data-generation mechanism whose value is mediated by the downstream SR learner.

6. Position within the rectified-flow literature

RFDM belongs to a broader family of methods that repurpose rectified flow beyond unconditional generation. "FlowIE" constructs a linear many-to-one transport mapping for image enhancement and emphasizes that rectification straightens trajectories of probability transfer, enabling fewer than 5 inference steps (Zhu et al., 2024). "IR-Flow" uses rectified flow to bridge discriminative and generative restoration by constructing multilevel data distribution flows and learning cumulative velocity fields across varying degradation levels (Fan et al., 21 Apr 2026). "WaterFlow" uses rectified flow for underwater salient object detection and explicitly injects physical degradation cues into feature hierarchies rather than removing them (Li et al., 14 Oct 2025).

Against that background, RFDM in super-resolution is distinctive because it models real-world degradation itself. Rather than transporting degraded inputs to clean outputs, it transports DT-LR images to the real-LR distribution so that pseudo-paired supervision can be constructed for an off-the-shelf SR network (Zhou et al., 10 Aug 2025). This suggests a modular interpretation of rectified flow in low-level vision: the same straight-path transport formalism can be used for restoration, synthesis, feature conditioning, or degradation modeling, depending on what distributions are chosen as endpoints.

The contrast with "TumorGen: Boundary-Aware Tumor-Mask Synthesis with Rectified Flow Matching" is also clarifying. The TumorGen summary explicitly states that there is no explicit mention of a "Rectified Flow Degradation Module (RFDM)" in that paper; rectification there is achieved through the integration of the Spatial-Constraint Vector Field Estimator and the VAE-Guided Mask Refiner within a rectified-flow framework (Liu et al., 30 May 2025). Thus, RFDM should not be treated as a universal component of rectified-flow systems, but as a task-specific design pattern that emerges when degradation itself is the object of transport modeling.