Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

An Empirical Study of Sampling Hyperparameters in Diffusion-Based Super-Resolution

Yudhistira Arief Wibowo

TL;DR

The paper investigates how conditioning hyperparameters affect diffusion-based single-image super-resolution, focusing on Diffusion Posterior Sampling (DPS) and Manifold Constrained Gradient (MCG) as conditioning schemes. Using a fixed pretrained diffusion prior on FFHQ, it shows that the conditioning step size largely governs performance, with optimal values around 2.0–3.0, while the number of diffusion steps plays a secondary role. DPS consistently yields better overall reconstructions than the baseline and MCG, which improves pixel-level fidelity but can harm perceptual quality. The work highlights the practical importance of tuning conditioning parameters in diffusion-based inverse problems and notes dataset-specific considerations for generalization.

Abstract

Diffusion models have shown strong potential for solving inverse problems such as single-image super-resolution, where a high-resolution image is recovered from a low-resolution observation using a pretrained unconditional prior. Conditioning methods, including Diffusion Posterior Sampling (DPS) and Manifold Constrained Gradient (MCG), can substantially improve reconstruction quality, but they introduce additional hyperparameters that require careful tuning. In this work, we conduct an empirical ablation study on FFHQ super-resolution to identify the dominant factors affecting performance when applying conditioning to pretrained diffusion models, and show that the conditioning step size has a significantly greater impact than the diffusion step count, with step sizes in the range of [2.0, 3.0] yielding the best overall performance in our experiments.

An Empirical Study of Sampling Hyperparameters in Diffusion-Based Super-Resolution

TL;DR

The paper investigates how conditioning hyperparameters affect diffusion-based single-image super-resolution, focusing on Diffusion Posterior Sampling (DPS) and Manifold Constrained Gradient (MCG) as conditioning schemes. Using a fixed pretrained diffusion prior on FFHQ, it shows that the conditioning step size largely governs performance, with optimal values around 2.0–3.0, while the number of diffusion steps plays a secondary role. DPS consistently yields better overall reconstructions than the baseline and MCG, which improves pixel-level fidelity but can harm perceptual quality. The work highlights the practical importance of tuning conditioning parameters in diffusion-based inverse problems and notes dataset-specific considerations for generalization.

Abstract

Diffusion models have shown strong potential for solving inverse problems such as single-image super-resolution, where a high-resolution image is recovered from a low-resolution observation using a pretrained unconditional prior. Conditioning methods, including Diffusion Posterior Sampling (DPS) and Manifold Constrained Gradient (MCG), can substantially improve reconstruction quality, but they introduce additional hyperparameters that require careful tuning. In this work, we conduct an empirical ablation study on FFHQ super-resolution to identify the dominant factors affecting performance when applying conditioning to pretrained diffusion models, and show that the conditioning step size has a significantly greater impact than the diffusion step count, with step sizes in the range of [2.0, 3.0] yielding the best overall performance in our experiments.

Paper Structure

This paper contains 14 sections, 4 equations, 5 figures.

Table of Contents

  1. Introduction
  2. Methodology
  3. Score-Based Diffusion Model as Image Priors
  4. Network Architecture.
  5. Usage as an Image Prior.
  6. Manifold Constrained Gradient
  7. Diffusion Posterior Sampling
  8. Experiment
  9. Setup and Configuration
  10. Experimental Results
  11. Quantitative Analysis.
  12. Qualitative Analysis.
  13. Conclusion
  14. Supplementary Tables

Figures (5)

  • Figure 1: Architecture of pretrained diffusion model
  • Figure 2: Geometric interpretation of conditioning strategies in diffusion-based inverse problems dps
  • Figure 3: Performance of models under various step sizes ($\zeta_t$) with step count 1000. Step size of 0 signifies no DPS application on the diffusion model.
  • Figure 4: Performance of models under various step count with $\zeta_t$ of 1.0.
  • Figure 5: Result of several test images. Ground Truth (GT) images are sourced from a publicly available image set.