Closed congdm closed 3 months ago
We suggest learning the variance through the variational bound during training. As noted in IDDPM, this helps the diffusion model learn more modes, creating more peaks in probability distribution. For text-to-image tasks, this enhances the diversity of the generated images. For more details, please refer to paper.
Hi, I'm currently writing training script for ipex based your example training script, I've noticed that in the training_losses function, there is an additional vb_terms when calculating the loss:
I've seen that this vb_terms was added to the final loss, but I don't understand the significance of this "variational bound". If I ignore it like this in my training script, what would be the impact?