Introduction
Diffusion models have emerged as a powerful class of generative models, particularly for tasks like Image Super-Resolution (SR). However, their stochastic nature introduces a "sampling lottery": the quality of the reconstructed high-resolution image depends heavily on the initial random noise vector. This instability often results in inconsistent textures or artifacts, posing a significant challenge for practical applications where reliability is paramount. While most research focuses on retraining models to improve performance, this work explores a different path: identifying an "approximately optimal" starting noise vector to consistently produce high-quality results from pre-trained models without additional training.
Methodology: Finding the "Golden" Starting Point
The researchers propose shifting the perspective from Stochastic Differential Equations (SDEs) to deterministic Diffusion Ordinary Differential Equations (ODEs). In a Diffusion ODE, the path from the initial noise ($x_T$, the boundary condition) to the final image ($x_0$) is fixed. The goal is to find a specific boundary condition, $\tilde{x}_T$, that yields the best perceptual quality across various inputs.
The method relies on three key steps:
- The Independence Discovery: Bayesian analysis suggests that the "optimal" starting noise is largely independent of the specific low-resolution input image.
- Monte Carlo Exploration: A set of noise vectors is randomly sampled and tested against a small reference set of high-resolution/low-resolution image pairs.
- Perceptual Selection: The noise vector that consistently minimizes the Learned Perceptual Image Patch Similarity (LPIPS) metric across the reference set is selected as the "approximately optimal" boundary condition ($\tilde{x}_T$).
Experimental Setup
The method was evaluated using standard super-resolution benchmarks. The baseline diffusion models were trained on the DF2k dataset (a combination of DIV2K and Flickr2K). Testing was conducted on DIV2k-test, Urban100 (rich in architectural details), and BSD100. Additionally, the approach was applied to StableSR, demonstrating its effectiveness on real-world degradations using pre-trained Stable Diffusion weights.
Key Results and Insights
The use of an optimal boundary condition provides a significant performance boost for existing diffusion samplers. Notably, it allows faster samplers to outperform slower, stochastic ones. For instance, a 100-step DDIM sampler using the optimal BC achieved an LPIPS score of 0.1032 on DIV2k-test, surpassing the 1000-step DDPM sampler (0.1075) while being ten times faster.
Beyond quantitative gains, the method ensures stability. By removing the randomness of the initial seed, the model produces steady, predictable, and high-fidelity images with fewer hallucinated artifacts. This "inference-side" optimization requires no changes to the model's weights, making it a "free" upgrade for pre-trained diffusion systems.
Conclusion and Future Directions
While highly effective, the current method is constrained by image resolution; an optimal noise vector found for a specific resolution (e.g., 256x256) cannot be directly applied to other sizes. Furthermore, the initial search for $\tilde{x}_T$ involves a one-time computational cost. Future research aims to develop "resolution-arbitrary" sampling methods and extend this optimal boundary logic to other image restoration tasks such as colorization and low-light enhancement.
Summary updated