Network Rank (Dim) and Network Alpha 3

When changed the Network Rank (Dim) and Network Alpha settings for LoRA training, found that the resulting LoRAs could no longer reproduce certain features that had been clearly expressed by LoRAs trained with my previous settings.

The original settings used in kohya were:

• Network Rank (Dimension): 8

• Network Alpha: 4

The newer settings were:

• Network Rank (Dimension): 64

• Network Alpha: 32–56

Specifically, the LoRAs trained with the newer settings produced less pronounced and less distinctive effects.

This is only a hypothesis, but suspect that using an Alpha-to-Rank ratio below 0.50 may have contributed to the development of a more distinctive style in the LoRAs trained with previous settings.