Trained with about 1000 tagged images of different characters, objects, styles. It performs best in photorealistic, portrait scenarios and creates very convincing (non standard) faces and poses. It's also compatible to a lot of LORAs I tested it with.
It often tends to wander of into the NSFW abyss, but with proper prompting this isn't an issue. Just be aware it can be naughty.
Currently, work in progress - and to be continued (after a short break). I will add more tagged images to the lot using Ollama with LLaVa 1.6 (which should allow detailed tagging NSFW).
Description
Initial release.
FAQ
Comments (4)
Not quite sure if i am doing something wrong here but I found this model incredibly rigid and sadly very unresponsive to prompts.
In spite of clear and precise prompts it will insist on returning very standardised imagery both in subject and backgrounds (i.e. a twenty ish brunette outdoors).
trying a second run now to exclude the possibility of memory leakage.
You're right. It's very stiff at times. With an adjusted CFG value you will see a difference, but I will try to broaden the training datasets a bit more.
@redhome On the other hand it generates great still-life images. Actually the best output of the 7 models I tested.
It is a decent expressive, high-quality and flexible photorealistic model. Like it. Respect!