Spectral Guidance for Flexible and Efficient Control of Diffusion Models

Gabriel Moreira, Manuel Marques, João Paulo Costeira, Chenyan Xiong

arXiv:2605.28900 · 2026-05-29 공개 · arXiv · PDF

diffusion-models self-supervised-learning cifar-10 phase-transition clip-embeddings conditional-control sampling-trajectory spatial-control

Abstract

We introduce Spectral Guidance, a framework for controlling diffusion models by leveraging the intrinsic geometry of the generative process. As data is progressively corrupted by noise, only a small number of features remain informative for control. We characterize them as the singular functions of a conditional expectation operator and show that they can be learned via a self-supervised objective. Once recovered, this basis enables the projection of arbitrary guidance signals, such as labels, CLIP embeddings, or masks, directly onto the sampling trajectory. This approach allows for stable, high-fidelity control without retraining or denoiser backpropagation during sampling. Empirically, we improve conditional accuracy on CIFAR-10 by 37 percentage points over the strongest training-free baseline while offering $4\times$ faster sampling. Moreover, the same representations that support label and CLIP guidance also enable spatial control, such as mask-based guidance, without auxiliary models. Finally, our framework reveals a phase transition in the generative process, pinpointing the optimal time window for effective guidance.

한국어 요약

📋 한 줄 요약

**[Diffusion / Training-Free Control]** Spectral Guidance가 conditional expectation operator의 singular function을 self-supervise로 학습 — label·CLIP·mask를 sampling trajectory에 projection, CIFAR-10에서 conditional accuracy +37pp·4× faster sampling, 생성 과정의 phase transition 발견.

🎯 핵심 기여도

💡 핵심 아이디어

Diffusion 제어는 생성 과정 시간에 따라 정보적 feature가 collapse하는 intrinsic geometry를 따르며, conditional expectation operator의 singular function이라는 작은 basis만 학습하면 label·CLIP·mask 등 다양 guidance를 trajectory에 projection으로 통합·가속할 수 있고, 효과적 guidance의 optimal time window가 phase transition으로 드러난다.

🔬 기술적 접근법

📊 주요 결과

💭 의의 및 한계

**의의**: Diffusion 제어를 intrinsic geometry로 재정초, 다양 guidance 신호 통합 처리·spatial control까지 auxiliary 모델 없이 확장, phase transition이 guidance scheduling의 원리적 가이드. **한계**: CIFAR-10 중심으로 high-resolution·복잡 conditioning 일반화는 후속, singular function basis 학습의 scaling 비용, 매우 fine-grained conditioning에서의 정확도는 추가 검증.

🚀 실용적 활용