During LoRA training, overfitting means the model learns the training images too specifically instead of producing reliable results beyond them. Sample prompts expose this: compare familiar training-like scenes with new environments, then inspect which checkpoint behaves best. section 13.3
Use this practical check before letting a training job run for hours:
- Select the correct dataset; a wrong dataset can waste a multi-hour training job. section 13.3
- Add a default caption containing the trigger word as a fallback, and leave Caption Dropout Rate at 0.0; its benefit against overfitting is not established. section 13.3
- Run sample prompts periodically, matching Sample Every to Save Every, and use the results to compare checkpoints. section 13.3
Prompts that reveal memorization
Start with familiar scenes such as “Keats the dog sitting in a fancy chair looking at the viewer” and “A photo of Keats the dog standing on a lakeside beach looking into the distance.” Add prompts placing the subject in novel environments to test generalization. section 13.3
Figure 13.14 Sample prompts are very helpful when debugging a training job.

Inspect the sample-prompt setup used to debug a training job and judge whether the model has overfit its data. section 13.3
































