Reproduce “Bridging Fidelity-Reality with Controllable One-Step Diffusion for 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
Hao Chen, Junyang Chen, Jinshan Pan, Jiangxin Dong
CODSR, a Controllable One-Step Diffusion Network for Super-Resolution, addresses fidelity loss, inadequate generative prior activation, and text-image misalignment in one-step diffusion methods. The framework achieves a strong balance between structural fidelity and perceptual quality, demonstrating superior perceptual metrics like MUSIQ and NIQE compared to existing one-step and some full-step diffusion SR methods on real-world datasets.
Key insights
- Adaptive noise injection, guided by image gradients, allows for a region-discriminative activation of generative priors, improving perceptual quality without introducing artifacts in smooth areas.
- Modulating U-Net's intermediate features with uncompressed low-quality image information enhances fidelity more effectively than direct latent modulation, preserving the latent space's integrity.
- Explicit spatial alignment of text embeddings with image semantic regions, enabled by external segmentation models, provides more precise and effective textual guidance for super-resolution.
- Efficient one-step super-resolution can achieve a robust balance between structural fidelity and perceptual realism by carefully addressing information loss and generative prior activation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.