I've created a set of training images for StableDiffusion 2.1, as I couldnt find any classification images in 768x768.
Do note, that this is by no means a solved problem, and I am just testing out different class image sets and prompt combinations.
Refer to the table below, and based on which set of images you choose, select the prompt you want accordingly. Replace <instance> with the name of your instance.
| Folder | Class Prompt | Negative Prompt | Instance Prompt to use | Remarks |
|---|---|---|---|---|
person |
person |
None | <instance> person |
Your basic, gender-neutral class images. Mileage may vary (training this on my face, I've gotten very quite inaccurate results frequently) |
photo of a person |
photo of a person |
ugly, deformed |
photo of <instance> person |
Better results for photographic outputs, but feels like it doesnt retain faces that well, and different art styles are comparatively more difficult to achieve. |