search

hyn2028 / tpdm

Official code for "Improving 3D Imaging with Pre-Trained Perpendicular 2D Diffusion Models" (TPDM)
47 stars 4 forks source link

Training Model Tips #8

Open aribeiro05 opened 6 months ago

aribeiro05 commented 6 months ago

Hello! I'm trying to use tpdm for generating a 3D Unconditional Volume using two diffusion models trained with knee volumes. One of the models was trained with coronal slices and the other with axial slices. When I use those checkpoints, after 2000 sampling steps, the images (Volume) generated are completely noisy, almost white noise. Do you have any tips on what is going wrong? Is there a specific orientation of the slices used to train the models? My dataset was also small (10 volumes). Could that be one of the reasons for the issues? However, when I use the LDCT checkpoints, the TPDM is able to generate a 3D unconditional volume, and those models were also trained with a small dataset

Best regards.

hyn2028 commented 5 months ago

Hi, @aribeiro05.

Thank you for your interest in our research.

Does plain unconditional two-dimensional image generation for each of the primary and auxiliary models work well?