search

sail-sg / MDT

Masked Diffusion Transformer is the SOTA for image synthesis. (ICCV 2023)
Apache License 2.0
500 stars 35 forks source link

Training Settings #47

Open Yeez-lee opened 1 month ago

Yeez-lee commented 1 month ago

Hi, I am curious about your MDTv2_s2 settings. What is your lr_anneal_steps? Should I set it to 0? And how can I control the training steps? For example, how can I set the overall training steps to 400k? Thank you!

gasvn commented 1 month ago

We keep the same learning rate in our setting. And we just kill the running program when it reaches the required training step.

Yeez-lee commented 1 month ago

We keep the same learning rate in our setting. And we just kill the running program when it reaches the required training step.

Thank you! One more question is about cfg_scale. Do you always keep 3.8 for S, B, and XL models during inference?

gasvn commented 1 month ago

Yes, the value is obtained with MDTv1, and I keep using 3.8 for all models of MDTv2