.png){kind=icon}
.png){kind=icon}
Return
ImpRes: implicit residual diffusion models for image super-resolution
.png){kind=icon}
PRE
AI
DOI：10.1007/s00371-024-03717-4.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Single-image super-resolution (SISR) is a fundamental task in computer vision that faces challenges due to the loss of high-frequency information during image degradation, leading to a nonunique solution space. Current super-resolution (SR) methods often suffer from high-frequency texture distortion, excessive smoothing, and scale inconsistency. This study introduces an innovative implicit residual diffusion model (ImpRes) to address these issues. ImpRes enhances model convergence speed and high-frequency detail recovery through a residual prediction mechanism. It incorporates a Gaussian frequency decomposition module using Gaussian high-pass filters to emphasize high-frequency components, guiding accurate texture reconstruction. Additionally, ImpRes employs static implicit neural representation (SINR) during decoding to transform discrete image representations into a continuous local implicit image function, achieving precise content perception, flexible spatial sampling, and mitigating over-smoothing. Experimental results demonstrate that ImpRes outperforms most existing diffusion-based methods in terms of model convergence time, generation quality, and scale consistency, achieving a peak signal-to-noise ratio of 29.97 dB in 4 x face super-resolution tasks. Our implementation is available at: https://github.com/fineverse/ImpRes.
Keywords:
Diffusion models
Image super-resolution
Residual prediction
Implicit neural representation
Journal
IF:
2.9
Papers:
4.6K
Citations:
6.5K