ASASR: Adversarial Sobolev Alignment for Faithful Image Super-Resolution
Researchers propose ASASR, a framework for image super-resolution that addresses spectral misalignment between generative priors and natural image manifolds. The method uses Sobolev-induced Riemannian geometry and a parametric adversary grounded in the Riesz Representation Theorem to generate targeted negative samples, improving fidelity over standard approaches like Direct Preference Optimization.
Key facts
- ASASR addresses spectral misalignment in image super-resolution
- Framework uses Sobolev-induced Riemannian geometry
- Parametric adversary based on Riesz Representation Theorem
- Generates targeted negative samples for optimization
- Improves upon Direct Preference Optimization
- Focuses on faithful restoration over generative priors
- Colored noise transition kernel mirrors natural spectral decay
- Extensive experiments demonstrate effectiveness
Entities
—