6-22

Rearranged the number of repetitions in the dataset to train longer but use a lower percentage when merging.

Integrated IterComp and spo-lora.

The following trigger words were trained, but not all of them trigger easily:

Arab

Indian

mixed-race

Korean

Indonesian

Iranian

Latina

Russian

Mediterranean

Eurasian

caucasian

Persian (differentiated from Iranian by pupil color)

+Features

Try Training Real Style

It was originally a dataset for Chroma, but Illustrious v2 was released, and most of the mergers with sdxl real lineage models forget the concept.

My goal's to not lose concepts at all and be as realistic as possible, but the skin textures of this model aren't quite up to snuff and are not adequately trained.

Using Photograph as a trigger word, and some facial feature labeling e.g. (Russian Features), the filtered ffhq dataset is used here.

Then there are also some concepts that the original Illustrious lacked, but were found to be insufficiently trained as well.

At least this model doesn't do anything but realistic faces like those models whose posters are fabulous.

Two other models based on Illustrious v2.

Illustrious-cos

Illustrious mixed-CGI （It's a model merger）

It's time for the sdxl era to end, and for image generation, there are quite a few recent high quality models, especially those using gemini, that understand text quite a bit better than flux. But for pornographic content, look forward to Chroma, which, by the looks of it, is quickly taking shape as well.

I'm hoping to fill in some rare concepts for it by then, but am struggling to find suitable images (real photos). If you have a source for Rare Concepts images that are high enough resolution and not covered by Danbooru's tags, please be sure to let me know.