Core Concepts
A conditional diffusion model with probability flow sampling is proposed to efficiently generate high-quality super-resolution images while maintaining consistency with low-resolution inputs.
Abstract
The paper proposes an efficient conditional diffusion model with probability flow sampling (ECDP) for image super-resolution. The key highlights are:
-
Continuous-time Conditional Diffusion for Image Super-Resolution:
- The forward process gradually adds noise to high-resolution (HR) images while conditioning on low-resolution (LR) inputs.
- This process preserves the mean and variance of the HR images, making the model training easier and generation faster.
- The conditional score function is learned using a hybrid-parametrization denoiser network that combines the strengths of the ϵ-parametrization and the x0-parametrization.
-
Efficient Generation with Probability Flow Sampling:
- The learned conditional score function is used to efficiently generate super-resolution images via probability flow sampling, which is much faster than iterative sampling methods.
-
Image Quality Loss:
- An additional image quality loss is introduced to directly optimize the perceptual quality of the generated super-resolution images.
- This loss is computed efficiently using the adjoint method, without depending on the intermediate values of the probability flow ODE.
Extensive experiments on DIV2K, ImageNet, and CelebA datasets demonstrate that the proposed ECDP method achieves higher super-resolution quality than existing diffusion-based methods while having lower time consumption.
Stats
The forward process gradually adds noise to HR images while conditioning on LR inputs, preserving the mean and variance of the HR images.
The hybrid-parametrization denoiser network combines the strengths of the ϵ-parametrization and the x0-parametrization.
Probability flow sampling is used for efficient generation of super-resolution images.
An image quality loss is introduced to directly optimize the perceptual quality of the generated super-resolution images.
Quotes
"To reduce the time consumption, we design a continuous-time conditional diffusion model for image super-resolution, which enables the use of probability flow sampling for efficient generation."
"To improve the consistency of generated images, we propose a hybrid parametrization for the denoiser network, which interpolates between the data-predicting parametrization and the noise-predicting parametrization for different noise scales."
"Moreover, we design an image quality loss as a complement to the score matching loss of diffusion models, further improving the consistency and quality of super-resolution."