Why do the sampling results not correspond to the labels?

[9crAen]

Unconditional generation was used, the input to the model was clear images, and after the training was completed the training set plus random noise was used for testing, and it appeared that the test results could not be matched with the labels one-to-one.

[9crAen]

To correct this, what happens is that the sampling results correspond to other labels (e.g., image B), not labels (image A)