Wrose generated image compare to Convnextv2

[powermano]

This is using spark network and replace the final project layer of convnext to Unet decoder. The loss is 0.16 roughly and the vis results is as following:

This using convnextv2 , the loss is 0.17 roughly

I think using unet decoder will let model learn some shortcuts for generate the images, as the training loss is almost the same.