Coloring the Noise: Adversarial Sobolev Alignment for Faithful Image Super Resolution

Generative priors in Image Super-Resolution (SR) often compromise faithful restoration, we attribute this limitation to a fundamental spectral misalignment between isotropic objectives and the intrinsic natural image manifold. While Direct Preference Optimization offers a path to alignment, its reliance on spectrally flat Gaussian noise fails to distinguish authentic high-frequency details from hallucinations. To bridge this geometric gap, we propose ASASR, a theoretically grounded framework that recasts the generative flow into a Sobolev-induced Riemannian geometry by explicitly coloring the noise transition kernel to mirror natural spectral decay. Driving this geometric alignment, we integrate a parametric adversary grounded in the Riesz Representation Theorem, which synthesizes targeted negative samples equivalent to worst-case Sobolev gradients to direct optimization along the tangent space of plausible structural failures. Extensive evaluations demonstrate that ASASR outperforms leading generative baselines, particularly in preserving spectral consistency and structural fidelity, offering a robust solution that effectively mitigates artifacts.

📋 한 줄 요약

**[Image Super-Resolution / Sobolev DPO]** ASASR이 generative SR의 spectral misalignment를 noise kernel coloring으로 해소 — Sobolev-induced Riemannian geometry + Riesz 적용 parametric adversary가 worst-case Sobolev gradient 부정 샘플 생성, hallucination 완화·spectral·structural fidelity 우월.

🎯 핵심 기여도

Image Super-Resolution(SR)의 generative prior가 faithful restoration을 compromise하는 한계를 isotropic objective와 intrinsic natural image manifold 사이의 fundamental spectral misalignment에 기인한다고 진단.
Direct Preference Optimization(DPO)이 alignment 경로를 제공하지만, spectrally flat Gaussian noise에 의존해 authentic high-frequency detail과 hallucination을 구분하지 못함을 지적.
ASASR 제안 — theoretically grounded 프레임워크로 generative flow를 Sobolev-induced Riemannian geometry로 recast: noise transition kernel을 explicit하게 coloring해 natural spectral decay를 모방.
Geometric alignment를 추진하기 위해 Riesz Representation Theorem 기반 parametric adversary 통합 — worst-case Sobolev gradient에 해당하는 targeted negative sample을 합성, plausible structural failure의 tangent space를 따라 최적화 진행.

💡 핵심 아이디어

Generative SR의 hallucination 문제는 isotropic objective(spectrally flat noise)와 natural image의 spectral decay 간 geometric 부정합이 근원이며, noise kernel을 natural spectral decay에 맞게 coloring하고 Riesz 정리 기반 parametric adversary로 worst-case Sobolev gradient 부정 샘플을 합성하면 spectral·structural fidelity를 동시에 보존하는 alignment가 가능하다.

🔬 기술적 접근법

**방법론**: ASASR — Adversarial Sobolev Alignment for SR.
**핵심 기법**: (1) Generative flow를 Sobolev-induced Riemannian geometry로 recast, (2) Noise transition kernel을 natural spectral decay에 맞춰 explicit coloring, (3) Riesz Representation Theorem 기반 parametric adversary, (4) Worst-case Sobolev gradient에 equivalent한 targeted negative sample 합성, (5) Plausible structural failure의 tangent space 방향으로 최적화 유도.

📊 주요 결과

ASASR이 leading generative baseline을 outperform.
Spectral consistency·structural fidelity 보존에서 특히 우수.
Hallucination·artifact 완화 효과 입증.
광범위 평가로 robust 솔루션 확인.

💭 의의 및 한계

**의의**: SR의 spectral misalignment 문제를 geometric·이론적으로 정식화, DPO를 Sobolev 기하로 일반화한 첫 사례 중 하나, hallucination vs detail의 trade-off에 원리적 접근. **한계**: Sobolev geometry·Riesz 기반 분석의 구현 복잡성, 실시간 SR 적용 시 비용, 다양 degradation type(blur·noise mix)으로 일반화 추가 검증.

🚀 실용적 활용

Real-world image SR과 photo restoration.
의료·위성 imagery의 faithful 복원.
Generative restoration의 spectral 일반화 기법.