Dancing where angels fear to tread, I've generated a list of racial groups including mixed races, for example: Chinese-Malay, Afro-Latino, etc.
This is to assist image generation models apply random keywords within an prompt. It should be saved as a .txt file in the wildcards directory if generating images locally in Fooocus, and wildcards_official for Ruined Fooocus. For A1111: This is a Wildcard collection, it requires an additional extension in Automatic 1111 to work. on that platform.
Usage: Simply put double underscores around the filename prefix. i.e.(__races__ in the prompt. For example "Close up shot of a beautiful 18 year old __races__ woman"
This will then cause one of the items in the list to be inserted into the prompt. For example "Close up shot of a beautiful 18 year old Kurdish woman", next prompt in the batch could be "Close up shot of a beautiful 18 year old Asian woman".
Keep in mind that the checkpoint or lora may not recognize the style and will produce a 'default' response, and images generated are highly dependent on the training material; 'afro-caribbean-indian' may not be recognized in your favorite anime checkpoint.
Comments, suggestions, amendments and inclusions most welcome in the comments.
Description
This version substantially increases the number of racial groups, with some cross-over into ethnic groups.
It has both 'generic' groups such as 'Asian' as well as specific groups such as Hmong, Japanese etc.
FAQ
Comments (8)
If you want a particular missing race added, let me know here in the comments.
The demo images were built using Ruined Fooocus, and unfortunately the build metadata isn't picked up automatically. It is embedded in the image file though.
Base mode was JuggernautXL8, with skin_texture_v2.safetensors Lora at .75
Thanks for putting this together, I love the idea of it. However I think most checkpoints can't handle things like this well (as you mentioned in the model description). Do you recommend any specific checkpoints that handle diverse races exceptionally well?
@fauxrealistic I've been using JuggernautXL with my testing and it actually gets about 95% of them, which is pretty amazing. When it fails, they tend to look like the descendants of Charles II of Spain.
I intend to expand my testing and see what others get good results. If I find any, I'll put them in the Suggested Resources section.
@fauxrealistic Now tested with newrealityxl_20 with good (better than JuggernautXL) results. Still some 'generic' women popping in, but that is to be expected.
@swedishViking ha! funny failure note. and thank you for this helpful info with models/suggested resources.
baltic, basque, bavarian, finno-ugric, flemish, frisian, gaelic, mediterranean, szekely
assyrian, kalasha, nuristani, rohingya, sentinelese
apache, comanche, nahua, pueblo, sioux
herero, hutu, malagasy, tutsi
Many thanks. Expect to see them in the next version.
At the first glance I thought it was a Jagged Alliance portrait lora.