LipsFlower constrains flow matching to stabilize generative image restoration
Original title: Lipschitz-Controlled Flow Matching for Robust Image Restoration
LipHDiT uses spectral normalization, bounded modulation and finite-radius gain penalties, tested on blind face restoration.
The paper introduces LipsFlower to stop flow-matching restorers from amplifying small input perturbations. It derives stability bounds for Flow-from-Y and observation-conditioned posterior sampling, then uses LipHDiT, a Lipschitz-aware hourglass diffusion transformer, as the vector field. On blind face restoration, the authors report lower perturbation amplification with restoration quality competitive with the evaluated baselines.
Unconfirmed: Robustness and quality gains are reported by the authors against their chosen baselines only.
- LipHDiT adds stabilized approximate spectral normalization, bounded modulation and convex residual mixing.
- Adversarial finite-radius gain regularization penalizes sensitivity to state and observation.
- Both Flow-from-Y and observation-conditioned sampling are evaluated on blind face restoration.
- AI models
- LipsFlowerLipHDiT
From the original article
1 min read at huggingface.co
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.
By Sedjro Salomon Hotegni, Konstantin Sonntag, Paul-Christian Bürkner · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.