Loss Changes

[Philipuss1]

Added two new loss functions and also made a few small changes

[illtellyoulater]

@Philipuss1 hi, may I ask what those two functions do and how they would improve the generation? Ty!

[Philipuss1]

@illtellyoulater Well, you can experiment with them in the latest version of Majesty Diffusion afaik, but in general, mean loss makes the picture more white and kind of make it look like a sketch, while the var loss is the same as mean loss, but black instead of white. It won't improve your image generation, but if you seek a specific effect on the image, these two loss functions can help

[illtellyoulater]

Thanks! But how do I call those functions? Should I include something in the diffusion schedule?

[Philipuss1]

@illtellyoulater You can just change the mean_scale and var_scale variables

[Dango233]

Added two new loss functions and also made a few small changes

Thank you for making this PR! I did a slightly different implementation - so now this loss works directly on latent space., and having a threholding parameter to it.

var_index, var_range = ([0]*1000,0.5)
mean_index, mean_range = ([0]*400+[0]*1000, 0.75)

var_index is the var_scale on different timesteps, var_range defines when the var_loss kick in - only when var > var_range, the var_loss will kick in.

[Philipuss1]

@Dango233 Ooh this is great! Much better than what I did!
Also, should I close the PR? I feel like it's pointless keeping it open anymore 🤔