yuanzhi-zhu
commented
1 year ago - Our approach utilizes pre-trained diffusion models as generative denoiser prior, while SR3 is a model trained directly for SR. In other words, SR3 is a conditional diffusion model with the low-resolution image as condition input. To know more details about our approach, motivation and method, please have a look at this slide.
- In order to output 512*512 images, you just need to employ unconditional diffusion models that are trained on images with size 512*512.
Hello, I am a beginner and English is not my native language, please pardon any incorrect descriptions. I have a few questions. Firstly, as far as I know, Google's SR3 super-resolution model also applies diffusion models to image restoration. I would like to know the differences between this model and SR3, as well as the overall architecture. What are the core innovations? Secondly, what efforts do I need to make if I want to input 512*512 images? Thank you very much for your answers!