LPNSR: Optimal Noise-Guided Diffusion Image Super-Resolution Via Learnable Noise Prediction
Summary: arXiv:2603.21045v5 Announce Type: replace-cross
Abstract: Diffusion-based image super-resolution (SR) aims to reconstruct high-resolution (HR) images from low-resolution (LR) observations. However, the inherent randomness injected during the reverse diffusion process causes the performance of diffusion-based SR models to vary significantly across different sampling runs, particularly when the sampling trajectory is compressed into a limited number of steps. A critical yet underexplored question is: what is the optimal noise to inject at each intermediate diffusion step?
In this paper, we establish a theoretical framework that derives the closed-form analytical solution for optimal intermediate noise in diffusion models from a maximum likelihood estimation perspective, revealing a consistent conditional dependence structure that generalizes across diffusion paradigms. We instantiate this framework under the residual-shifting diffusion paradigm and accordingly design an LR-guided multi-input-aware noise predictor to replace random Gaussian noise.
Key Contributions
- Theoretical Framework: Establishes a closed-form analytical solution for optimal noise in diffusion models.
- Noise Prediction: Introduces an LR-guided multi-input-aware noise predictor to replace conventional random Gaussian noise.
- Mitigation of Initialization Bias: Utilizes a high-quality pre-upsampling network to enhance model performance.
- Compact Trajectory: Implements a 4-step trajectory that allows for end-to-end optimization, which is typically computationally prohibitive in longer trajectories.
Experimental Results
Extensive experiments conducted reveal that LPNSR achieves state-of-the-art perceptual performance across various datasets, including both synthetic and real-world images. Notably, the method does not rely on large-scale text-to-image priors, which often complicate the training of models in related domains.
By focusing on the injection of optimal noise at each diffusion step, LPNSR enhances the consistency and reliability of image super-resolution results. The research highlights the significance of noise management in diffusion processes, paving the way for future advancements in the field.
Conclusion
LPNSR represents a significant advancement in the domain of image super-resolution by addressing critical challenges associated with noise injection during the diffusion process. The findings contribute to a deeper understanding of noise dynamics in diffusion models and offer a robust framework for developing high-performance SR solutions.
Source Code
The source code for LPNSR can be accessed at https://github.com/Faze-Hsw/LPNSR.