How to compare style LoRA checkpoints before training again
Before starting a new round of style LoRA training, select early and late checkpoints and compare them using the same prompts, seeds, base model, resolution, and sampling settings, at both commonly used and lower LoRA strengths, and include a version with LoRA disabled. Then test with prompts for subjects missing from the dataset to check whether only the style was learned without copying objects or compositions from the training images. Finally, validate the best checkpoint and strength with multiple seeds, and record style match, prompt adherence, and artifacts labeled by checkpoint, optimizer step count, and LoRA strength.