Title: A Framework for Low-Level Perceptual Refinement in Unconditional Diffusion Models

Abstract:

While unconditional diffusion models provide robust generative priors, the challenge of directing them to produce aesthetically superior results remains largely under-investigated. This study demonstrates that h-space patching, currently the leading approach for training-free diffusion editing, is fundamentally inadequate for the global, low-level transformations necessary for aesthetic and perceptual enhancement. To address this, we propose a new, generalized framework for editing images within unconditional diffusion models that requires no explicit training. Our method functions as an inference-time mechanism that manipulates low-level features by isolating degradation concept vectors. By integrating bottleneck patching with classifier-free guidance, the approach steers the sampling process away from the degraded data manifold. This results in consistently higher-quality images, achieved without the need for any model retraining.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC