NTH

Lipschitz-Controlled Flow Matching for Robust Image Restoration

AuthorsSedjro Salomon Hotegni, Konstantin Sonntag, Paul-Christian Bürkner, Sebastian Peitz

AffiliationsPaderborn University · TU Dortmund University

November 16, 2026 2 min read
Watch on YouTube
The one-line take

LipsFlower aims to make flow-matching image restorations less sensitive to small input changes without sacrificing perceptual quality.

Key results

3000
CelebA-Test evaluation pairs

Shared test set for blind face restoration.

33%
Maximum gain reduction

Flow-from-Y under RMS-constrained SPSA; maximum gain falls from 1.74 to 1.16.

12%
CVaR95 gain reduction

Flow-from-Y under RMS-constrained SPSA; worst-five-percent average gain falls from 0.849 to 0.747.

26%
LPIPS drift reduction

Flow-from-Y under RMS-constrained SPSA.

0.31
PSNR decrease

Flow-from-Y restoration quality; PSNR changes from 25.784 to 25.475 dB.

What the paper found

LipsFlower targets a weakness in generative image restoration: small changes to a degraded image can produce large changes in the reconstruction. Its stability analysis distinguishes two pathways—perturbations entering through the initial state in Flow-from-Y, and perturbations affecting the velocity field throughout observation-conditioned sampling. To limit amplification, it introduces LipHDiT, an hourglass diffusion transformer with stabilized approximate spectral normalization, bounded modulation, and convex residual mixing, alongside adversarial finite-radius gain penalties at sampled flow states. On 3000 CelebA-Test face pairs at 512×512 resolution, the Flow-from-Y version reduces worst-case gain under an RMS-constrained SPSA attack from 1.74 to 1.16, a 33% decrease; its worst-five-percent average gain falls from 0.849 to 0.747, about 12%, and LPIPS drift drops about 26%. The robustness gain comes with a 0.31 dB PSNR decrease, from 25.784 to 25.475 dB. The conditional BayesFlow version lowers maximum amplification and perceptual drift but sacrifices more restoration quality. Comparisons include CodeFormer and OSDFace, which uses Stable Diffusion 2.1-base. The controls encourage stability but do not provide a global Lipschitz guarantee, and the evaluation is limited to blind face restoration.

Original abstract

Flow-matching models can produce high-quality image restorations, but their learned dynamics may amplify small input perturbations and make reconstructions unstable. We introduce LipsFlower to reduce this amplification. We derive stability bounds for Flow-from-$Y$, which initializes the ordinary differential equation from a noise-perturbed observation, as well as for observation-conditioned posterior sampling, which conditions the velocity field on the observation throughout integration. Based on this analysis, we develop LipHDiT, a Lipschitz-aware hourglass diffusion transformer that serves as the vector-field architecture of LipsFlower. LipHDiT incorporates stabilized approximate spectral normalization, bounded modulation, and convex residual mixing. We also use adversarial finite-radius gain regularization at sampled interpolation states to penalize excessive sensitivity to the state and, for conditional sampling, the observation. We test both formulations on blind face restoration, measuring adversarial robustness and the perceptual quality of the reconstructions. LipsFlower reduces perturbation amplification while keeping restoration quality competitive with the evaluated baselines. These results suggest that Lipschitz-aware flow matching can improve the stability of generative image restoration.

Read the original paper

More in Generative Models

Browse all 87 papers →
01Generative Model

dNaViT: Discrete Codes are General-Purpose Visual Representations

Xiaotong Li, Zhengrong Yue, Xu Huang, Jialv Zou, Minhao Jing, Wenhao Yan, Haodong Li, Hai Huang, Bo Zhao, Yan Feng, Peng Pei, Xunliang Cai

dNaViT turns images into reusable discrete codes designed to work across visual understanding, generation, and editing.

Read analysis