Why you call this model ddpm in your paper?
Opened this issue · 0 comments
Hi,
I can see from your code and the introduction of your paper that your method adds noise at a fixed time t during training and then directly predicts the original image for l1 loss constraint. In the inference stage, the noise is added to a fixed time t and the final result is predicted in one step.
I think it is more like a unet debluring model than ddpm. Because ddpm is multi-step iterative to predict noise, and the earliest work in 2015 is multi-step iterative to predict the original image. You use a fixed number of step to add and remove noise, and predict the original image in one step. Is this really a diffusion model?
Thank you very much for your work! This is not to question the validity and authenticity of your work, but to ask why it belongs to the diffusion model.