search

VisualComputingInstitute / diffusion-e2e-ft

Fine-Tuning Image-Conditional Diffusion Models is Easier than You Think. Accepted to WACV 2025 and NeurIPS AFM Workshop.
https://vision.rwth-aachen.de/diffusion-e2e-ft
326 stars 4 forks source link

The marigold depth model always produce "nan" #2

Closed hirotong closed 1 month ago

hirotong commented 1 month ago

Hi, Thanks for the great work! I found some problems with the marigold-e2e-ft-depth model that it always gives "nan" prediction. I checked everything is fine before calling the unet But I got all "nan" in the noise_pred. The default configuration is used and I also tried the online demo using the same model and it worked amazing.

Could you please help with the issue?

Thanks,

hirotong commented 1 month ago

Problem solved. Turned out to be some problems with the model weights. I re-download the weights and the problem disappears.