After you have finished training a model, we recommend creating 10 instance images
This will show you what the training retained and whether your model is overfit. You should then test your model with a few simple prompts.
If you notice that your images are including details from your original dataset that you did not want, you may be able to prompt them out. However, if testing the model with prompts does not produce the desired results, you will need to start again. In this case, we recommend referring to our tutorials on curating datasets and training models.
Another path you can try is to create a LoRA Composition with your new model to determine if adding concepts to the model can improve the output quality. Added a custom fine-tune to a composition is always recommended as it is an easy way to increase the fidelity of the model. For larger datasets, you can even try breaking up the dataset and training multiple models and blending them together!