Adaptive Lambda Scheduling in the DREAM Super-Resolution Model

Author

Kaiyue Wu

Date of Award

6-2-2025

Document Type

Thesis - SCU Access Only

Publisher

Santa Clara : Santa Clara University, 2025

Degree Name

Master of Science (MS)

Department

Computer Science and Engineering

First Advisor

Nam Ling

Abstract

Diffusion-based super-resolution (SR) models often suffer from a training–inference discrepancy: during training, they denoise images conditioned on ground-truth high-resolution targets, but during inference, they rely solely on self-generated predictions. The DREAM framework (Diffusion Rectification and Estimation-Adaptive Models) addresses this mismatch by blending model predictions with ground-truth supervision through a time-dependent weight, λ_t .

In this work, we investigate how the scheduling of λ_t affects SR quality. Specifically, we evaluate three λ_t scheduling strategies: a power-law schedule (used in the original DREAM), a cosine-based curve, and a sigmoid-shaped curve. Each defines a distinct way to balance reliance on ground truth versus the model’s own predictions throughout the diffusion process.

We conduct 16× super-resolution experiments on the CelebA-HQ dataset, measuring fidelity with PSNR and SSIM, and perceptual quality with LPIPS and FID. Results demonstrate that λ_t scheduling significantly influences the trade-off between pixel-level accuracy and perceptual realism. Among the tested strategies, cosine scheduling consistently provides the best balance. These findings highlight the critical role of adaptive λ_t design in improving the performance of diffusion-based SR systems.

Recommended Citation

Wu, Kaiyue, "Adaptive Lambda Scheduling in the DREAM Super-Resolution Model" (2025). Computer Science and Engineering Master's Theses. 48.
https://scholarcommons.scu.edu/cseng_mstr/48