Dreambooth Training SDXL Using Kohya

I will skip what SDXL is since I’ve already covered that in my <a href="https://medium.com/@yushantripleseven/dreambooth-sdxl-using-kohya-ss-on-vast-ai-10e1bfa26eed#98d5" rel="noopener">vast.ai guide</a> so I’ll just jump right in. If you want a more in-depth read about SDXL then I recommend <a href="https://medium.com/towards-data-science/the-arrival-of-sdxl-1-0-4e739d5cc6c7" rel="noopener">The Arrival of SDXL</a> by Ertuğrul Demir kohya_ss supports training for LoRA, Textual Inversion but this guide will just focus on the Dreambooth method. Much of the following still also applies to training on top of the older SD1.5 and SD2.x models. There are two ways to go about training the Dreambooth method: Token+class Method: Trains to associate the subject or concept with a specific token word (identifier)+class. i.e. <code>“ohwx person”</code> No need to provide captions. For example, if you use it to train a specific character, it is easy because you do not need to prepare a caption, but all features of the training dataset such as hairstyle, clothes, background, etc. are learned by being linked to the token word identifier, so when generating, there may be situations where you can’t change clothes easily with prompts. Caption Method: This involves creating a text file containing captions that describe each image in the dataset. For example, if you want to train a specific character, by describing the details of the image in the caption i.e.<code>“ohwx person wearing a white shirt, standing in the middle of a busy street, people in the background.”</code>, you can separate the subject from other elements, making it more precise. You can leave out the class word <code>“person”</code> from the caption and some people prefer the results this way. It is best to experiment and see which works best for you. <a href="https://medium.com/@yushantripleseven/dreambooth-training-sdxl-using-kohya-ss-windows-7d2491460608">Visit Now</a>

Dreambooth Training SDXL Using Kohya_SS (Windows)