What is your version of arcface_checkpoint.tar?

[ygtxr1997]

I replaced your 'arcface-checkpoint.tar' with arcface-100 pretrained on MS1M. After training SimSwap, I found that the identity similarity seems largely worse than your provided SimSwap-224.

Target, source, and swapped result are shown below:

[netrunner-exe]

I replaced your 'arcface-checkpoint.tar' with arcface-100 pretrained on MS1M. After training SimSwap, I found that the identity similarity seems largely worse than your provided SimSwap-224.

Target, source, and swapped result are shown below:

I think they took the checkpoint from this repository. It is possible that if the model was initially trained using arcface-100 pretrained on MS1M, then the result would be better (but this is not certain)

https://github.com/foamliu/InsightFace-v2/releases/tag/v1.0

[ygtxr1997]

I think they took the checkpoint from this repository. It is possible that if the model was initially trained using arcface-100 pretrained on MS1M, then the result would be better (but this is not certain)

https://github.com/foamliu/InsightFace-v2/releases/tag/v1.0

Thanks for your answer! I'm not sure if I mistakenly set the error backward propogation during training. I used the same model codes (generator and discriminator) but re-wrote the training codes using pytorch-lightning by myself, including dataloader, loss function, and so on. Maybe I need to check the training logic and receipe more carefully. I show my training losses below, can you find some possible mistakes in these metrics?

[ygtxr1997]

In my case, the mistaken face alignment scheme is probally responsible for the identity inconsistency issue. Closing.

[renmengyuan]

In my case, the mistaken face alignment scheme is probally responsible for the identity inconsistency issue. Closing.

Hi， what kind of face alignment scheme is better when you train the model?