Sometimes I had a strange training output with below settings:
Original model -> 87.5 coverage -> Batch 16/32/64 -> smooth/ssim -> Extended mask -> default kernel blur on masks
The preview after 50k iterations of batch size 64 still looks like A->A rather than A->B. Esp with ssim loss, I see this often.
Is there any tech reason behind this ?
I made sure I chose similar skull types and ensured features resemble closely on A & B faces.