Reading: https://rentry.org/RVC_making-models.
https://media.discordapp.net/attachments/1065253271540875335/1149644675653832744/modecollapse.png
It outlines mode collapse being an issue that is only resolved though a bit of luck and retraining. But would it not be possible to resume training from checkpoints before the mode collapse? If an eval g/total threshold is dropped below, it would just reload the last few saved checkpoints and continue training from there.