Plug-in Image Quality Control for Posterior Diffusion Super-Resolution

Dongkeun Oh

Keywords: super resolution, diffusion, posterior learning, KL divergence, finite difference, diffusion prior, inverse problem, numerical error

TL;DR: numerical error correction term for posterior learning is used for image based classifier guidance to enhance fidelity or peceptual quality of an image in the manner of plug and play.

Abstract:

Diffusion-based super-resolution (SR) has shown remarkable progress, mainly through prior-guided approaches that require explicit degradation models or semantic priors. While posterior diffusion SR avoids these assumptions by directly learning from LR–HR pairs, it still suffers from numerical errors during sampling and lacks plug-in mechanisms for quality control.
We introduce the first plug-in framework for posterior diffusion SR, enabling pretrained models to support controllable quality through marginal calibration, without retraining the core diffusion model. Our numerical analysis reveals that discretization errors are a key bottleneck in posterior SR. We prove that these errors can be equivalently expressed as gradients of KL divergence, unifying numerical error correction with image based classifier guidance. This provides a principled explanation of fidelity degradation and a new lens for posterior diffusion trajectories. In principle, these errors can be corrected to improve fidelity when reference supervision is available, offering a new theoretical understanding of posterior diffusion trajectories.
In real-world SR, we further show that our image based guidance offers a controllable trade-off between fidelity and perception, delivering perceptual sharpness competitive with state-of-the-art prior-based models.
Experiments confirm that our method consistently improves perceptual quality, while also validating the theoretical link between numerical errors and fidelity in posterior SR. These results position our work as a new direction for posterior diffusion models, bridging probabilistic analysis with practical deployment.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 14514