How does layer-wise feature matching help with discriminator and GAN training objective?

[Zephyr69]

The GAN_feat_loss seems to be calculating MAE between feature maps produced by the discriminator when it is doing forward pass using real and fake images.

How does this help with the training of discriminator and the whole GAN? I thought the discriminator could (and should) totally produce very different feature maps when forward passing real and fake images and still be effective, especially when using BCEWithLogitsLoss instead of MAE or MSE in discriminator losses. Wouldn't minimizing this loss lead to collapse?

Could someone explain this?