search

StanfordMIMI / DDM2

[ICLR2023] Official repository of DDM2: Self-Supervised Diffusion MRI Denoising with Generative Diffusion Models
140 stars 21 forks source link

Question about the beta schedule #31

Open Calvin-Pang opened 5 months ago

Calvin-Pang commented 5 months ago

Hi, thanks for publishing the code! I am trying to implement your code on other MRI modalities, whose noise levels may be less series than DWI images. However, I found the matched states in Stage 2 are quite small (around 0~50). Do you have any insights on how to use your method on less noisy images, or have you done these kinds of experiments? Thanks so much!

tiangexiang commented 5 months ago

Hi, thanks for your interest in our work! It is expected behavior that less noisy images should be matched with small state numbers (since they are more close to the clean reference). Do you encounter any problem when running Stage3? Can you observe visual improvements between Stage3 results and Stage1 results?

Calvin-Pang commented 5 months ago

Thanks for the reply! I do have tried to just use the few matched states for stage 3, but the generated images are not so good. Did you change the beta schedule? I noticed you use beta_0 = 5e-5, beta_T = 1e-2, which seems different from that of DDPM (I remember they are 1e-4 and 2e-2). Does it mean changing the starting and ending of the schedule can make stage 2&3 have larger matched states?

tiangexiang commented 5 months ago

In our original implementation, as you mentioned, we changed the beta scheduler to have large matched states (more generation steps). Yes, changing beta schedule in your case may help denoise less noisy images.

Calvin-Pang commented 5 months ago

Thanks so much! I also plan to change the beta schedule. Can you share any insights about how to modify the beta schedule based on your estimated noise level? Also, I am wondering generally how many steps you will match for your dataset (DWI images). Thanks!