Reproduce “GDPO-SR: Group Direct Preference Optimization for One-Step Generative Image Super-Resolution” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Qiaosi Yi, Shuai Li, Rongyuan Wu, Lingchen Sun, Zhengqiang Zhang, Lei Zhang

GDPO-SR presents a one-step generative image super-resolution framework that incorporates Group Direct Preference Optimization, enhancing generative capacity and visual quality while maintaining efficient inference. The method achieved leading performance, including the highest PSNR (28.18dB), SSIM (0.7839), MUSIQ (65.63), and CLIPIQA (0.7020) on the DRealSR dataset.

Key insights

  • Injecting controlled stochasticity into one-step diffusion models is crucial for enabling preference-based reinforcement learning, allowing for diverse sample generation from deterministic inputs.
  • An adaptive reward function, which dynamically weighs fidelity and perceptual quality based on image content characteristics, provides more effective guidance for generative ISR than static or single-metric rewards.
  • Group-based online preference optimization, by computing advantages relative to other generated samples, can effectively enhance both local detail reconstruction and overall visual quality in one-step ISR.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.