When training a style, you can choose between different intensity levels. Note that a higher intensity level does not always mean a better result. The best fit will depend on your dataset.
For a well-balanced, medium-sized dataset with a distinct style, any level should work well. This is demonstrated in the following images:
Prompt: "portrait of a woman with long purple hair standing in a city, 8k, high quality" using the same seeds
Prompt: “portrait of a man wearing a yellow sweater standing in a jungle, 8k, high quality” using the same seeds
Prompt: “a beautiful brutalist building by a lake in a beautiful landscape, 8k, high quality” using the same seeds
(Download Dataset)
As shown above, there is not much difference in quality between the levels. However, for intensity level 3, the backgrounds seem to have slightly more depth. So if your training images feature images with rich backdrops, a higher level can be beneficial, depending on how balanced your dataset is. A higher level can also be helpful if the style you are trying to capture is subtler.
Datasets that feature mostly the same subjects can be difficult to work with, requiring more complex prompt engineering. For such datasets, it is recommended to use intensity level 1, as level 3 would make them even less flexible.
Here are examples created using a model trained on a small dataset that did not have much variety:
Prompt: “a cute black owl bird, ruffled feathers” using the same seeds
(Download Dataset)
As you can see intensity level 3 depicts only very robot-looking figures, whereas the output from levels 1 & 2 has more owl-like ones.
intensity level 1 is good for:
smaller datasets → below 14 images
datasets that don’t feature a lot of variety, containing the same or very similar subjects
styles with high contrast, as those can look overbaked at a higher intensity level
intensity level 2 is good for:
most medium-sized datasets → 14 to 28 images
datasets that have some variety to them, like different subjects, different color schemes
intensity level 3 is good for:
medium to large datasets → above 28 images
datasets that have a lot of variety
images with a lot of background depth, as opposed to something that has more of a flat style
art styles that are rather subtle or have a softer contrast
Since datasets and styles can differ significantly and there is always some randomness to what the AI will pick up on, there is no one-size-fits-all solution, but these options should help you get good results.