Open kabachuha opened 12 months ago
I took a quick look at this paper, and its main idea is to add noise to adversarial samples and then use DDPM for denoising to obtain clean samples. If we want to use DiffBIR to generate clean samples, we need to ensure that the first-stage model is capable of removing this type of noise. We can start by verifying whether the current release of the first-stage model has this capability. If it doesn't, we can train a first-stage model to remove this noise and then use the second-stage model (without retraining it) to refine the details. Thank you for your consideration and suggestions!
Hi, all! Thanks for your work on such a project
I wonder if the image restoration capabilities can be extended to adversarial noise purification as well, similar to the year old Nvidia's DiffPure https://github.com/NVlabs/DiffPure
I'm certain such an application. which is able to both restore the images and clear them off adversarial noise, will have a profound impact on the industry