The variance tensor contains elements less than 0

[wanzhe45]

Thank you for your excellent work. I have successfully reproduced your method for obtaining pixel-wise uncertainty using Stable Diffusion; however, I have two questions regarding the implementation:

1. In the script you provided in the answer Can you give the visualization code for the uncertainty estimation for each pixel point?, should the line z_list.append(exp + torch.rand_like(exp) * dev) actually be z_list.append(exp + torch.randn_like(exp) * dev)?
2. Why are some elements in the saved variance tensor less than 0?