Can I not use real photos for anime loras even if I am training for an item and not a character?

Realistically how long until other kontext variant that has apache 2.0 license and nsfw combability is made (Chinese model). Do you think the model will benefit with training like Pony/chroma

>Chroma
>same prompt
>same resolution
>same cfg/steps/scheduler
>only changing seed
>Gen1 is genuinely incredible realism
>Gen2 is plastic
>Gen3 looks like tapestry woven from human hair
What the actual fuck is this model.

What should I have for NAG settings for imagegen?