>>105749188
it's cause people don't know how to train them. you have to have your pixels perfectly aligned and all same scaling (common is X4 or X8)
you can't just shove a bunch of randomly scaled pixel art into the training dataset, and you also have to disable bucketing or rescaling during training
yes there will be some minimal noise from the VAE but if you trained it well it means after rescale the result will be almost identical to the gen