Loss problem when training the conditional denoising U-net for diffusion

IanYeung / MGLD-VSR

Code for ECCV 2024 Paper "Motion-Guided Latent Diffusion for Temporally Consistent Real-world Video Super-resolution"

Other

95 stars 2 forks source link

Loss problem when training the conditional denoising U-net for diffusion #14

Open ssj9596 opened 4 months ago

ssj9596 commented 4 months ago

I am currently training a conditional denoising U-net for a diffusion model, but I am encountering an issue with the loss function. When the batch size is set to 1, the loss value remains at 1 and does not decrease. However, when I set the batch size to 4, the loss value becomes NaN. I would like to know which parameters I should adjust to resolve this issue. Thank you very much!

projectersMan commented 2 months ago

Hi , I have the same issue , do you solve the problem yet? would you like to share the method to solve it with me?