Conditional Guided Diffusion Probabilistic Models for Image Super-Resolution

Bin Xie, Hao Tang, Dawen Cai, Yan Yan

Keywords: Diffusion Models, Image Super-resolution

Abstract:

We propose a novel Conditional Guided Diffusion Probabilistic Model (CG-DPM) for image super-resolution. CG-DPM adopts diffusion models, which have strong abilities to generate diverse and photo-realistic images, through a stochastic iterative denoising process. The abilities can tackle the existing issue of over-smoothing artifacts in super-resolution tasks. The earlier work SR3 firstly uses diffusion models to conditional image generation for super-resolution. However, it simply upsamples the low-resolution images to the target resolution using bicubic interpolation as the conditional input, which cannot maximize the information of conditional images. In contrast, our CG-DPM involves conditional images in each different-scale level in the encoder so that the model can use the conditional images more effectively.
We also introduce a separate score-based likelihood model to guide the original diffusion model to obtain a score-based posterior model.
Moreover, since there is no analytic probabilistic formula to represent the likelihood probability for image super-resolution, we propose a novel scored-based loss function to train a separate guided network so that it can approximate the score-based likelihood probability.
We conduct experiments on image super-resolution tasks for human faces and natural images at different scaling factors.
CG-DPM achieves strong performance compared with existing methods.
Meanwhile, the proposed method can also be used on other tasks, and more experiments show that our method achieves competitive results on the medical image segmentation task.

Primary Area: generative models

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 7740