[Bug]: unexpected difference between latent image and latent conditioning image

[seniorsolt]

What happened?

I'm training finetune inpaint stable diffusion 1.5 with masking.

I think it's because there are two different modes for latent in SampleVAEDistribution.py from mgds:

def get_item(self, variation: int, index: int, requested_name: str = None) -> dict:
    distribution = self._get_previous_item(variation, self.in_name, index)

    if self.mode == 'sample':
        latent = distribution.sample()
    elif self.mode == 'mean':
        latent = distribution.mode()
    else:
        raise Exception('method not supported')

What did you expect would happen?

I expect to see difference between latent image and latent conditioning image only within masked area.

Relevant log output

No response

Output of pip freeze

No response