Lipschitz-Controlled Flow Matching for Robust Image Restoration
LipsFlower reduced sensitivity to input perturbations while keeping face-restoration quality competitive with evaluated baselines.
The researchers tested two flow-matching formulations on blind face restoration, measuring adversarial robustness and perceptual quality. One starts from a noise-perturbed observation; the other conditions on the observation throughout integration. Their LipHDiT architecture constrains how strongly the model can amplify changes, while an adversarial regularizer penalizes excessive sensitivity. The abstract reports improved stability but gives no numerical results.
HF Daily Papers' note
The researchers tested two flow-matching formulations on blind face restoration, measuring adversarial robustness and perceptual quality. One starts from a noise-perturbed observation; the other conditions on the observation throughout integration. Their LipHDiT architecture constrains how strongly the model can amplify changes, while an adversarial regularizer penalizes excessive sensitivity. The abstract reports improved stability but gives no numerical results.
HF Daily Papers' note
score 4