For context, the 224 and 512 pre-trained models released to public were trained with how many iterations?

[nonlin]

As the title asks. Bonus if the hardware and time it took to train to those iterations can be revealed as well.

[Fibonacci134]

As the title asks. Bonus if the hardware and time it took to train to those iterations can be revealed as well.

Its really a subjective thing. The model is light weight enough to be ran on a 4Gb GPU very comfortably and not cause a COOM; its more about how many iterations/epochs they trained on the weights. The 224 is an extremely accurate model, and the people complaining are usually comparing it to something like FA or DFL. Im just glad they chose to release all their work to the public even though they definitely could have profited from this. Some website actually was using Simswap as their SC and charging people for the service, they were shut down thankfully. I cant wait to see some of the training results people here produce on the 512 model!