Inference results of the overfitted model are noise

[kirilllzaitsev]

Having an overfitted 'base' part of the Imagen, I try to sample with it for 300 timesteps and get the following:

During training, it predicts the noise well, and MSE goes down steadily. Somewhat similar is the following issue lucidrains/imagen-pytorch#337.

Could you give any advice on the direction to investigate the issue?

[chiho-5]

For how many epochs did you train. I think I'm encountering the same problem

[sadath-12]

same problem

[swayampragnya-malla]

same problem

[kirilllzaitsev]

same problem

Try increasing the number of epochs (for instance, 1500 epochs (1 epoch = 1 gradient step) for a single image) and tuning the learning rate (starting with small values <1e-4)