There are a plethora of options for training Stable Diffusion models, each with their own advantages and disadvantages.
Most training methods can be used to train a singular concept such as a subject or a style, or multiple concepts simultaneously.
The choice of training method will be limited by your hardware, and will impact the quality of the images you produce.
In this guide, we'll cover the main training methods used today:
- Dreambooth: take existing models and incorporate new concepts into them
- EveryDream: think of this as training an entirely new Stable Diffusion, just a much smaller version
- LoRA: functions like dreambooth, but instead of changing the entire model, creates a small file external to the model, that you can use with models.
The custom checkpoint models you see on civitai.com? Most of them are trained with Dreambooth, or merges of other models (which may be merges themselves).
LoRA is the most popular training method today because it's the most accessible: it requires the least memory, trains the fastest out of these methods, and produces the smallest files (although they are more like 'add-ons' that you must use with existing checkpoint models)
The concept is what you want to teach the model. You can teach the model anything, as long as it can be represented in image-form.
Concepts generally fall under 2 categories: subjects and styles.
How many concepts in a single model?
There is no hard limit on the number of concepts you can train, but the more concepts you add, the more difficult it is to train them effectively and the longer training will take.
How can popular models have so many concepts?
It seems like popular models are trained on an absurd number of concepts!