Ok, but you're severely misrepresenting the importance of things. Base SDXL is a fine model. Base SDXL is going to be much better than a materially smaller model that you've retrained with "good data".