A practical stopping point for LoRA training

ostrisai · x · 2026-07-20

The post gives a practical rule of thumb for deciding when to stop a LoRA run: not at the very bottom, because some overfitting is desirable, but around 3k steps in the author’s setup.

The key idea is to add a validation loss curve to AI Toolkit so you can track how far training has moved past the model’s original validation loss without overfitting too much. The example uses a character LoRA trained on 19 images with 1 validation image, and the author says this makes the loss graph genuinely useful for judging when to stop.

Original post →

A practical stopping point for LoRA training

More from Multimodal