About 3x the training data
Does what it says on the tin. Be sure to use both the high-noise and low-noise LoRAs.