This fine-tune aims to find a balance to retain character familiarity and realistic results with life-like anatomy. I find it works the best with a blend of natural captions and booru tags as this fine-tune is trained on both, and natural captions help to make the model infer from more strictly realistic knowledge, while booru tags brings in anime-influences.

Recommended settings:

Sampler: radau_iia_2s, Scheduler: bong_tangent, Steps: 30. CFG:4

If using ComfyUI, this sampler is available with ClownsharKSampler. If you don't have this node installed then 2nd best option is res_2s sampler and beta57 scheduler.

It is trained on a procured dataset of 21,000 images with 33.3% NSFW, 33.3% cosplay and 33.3% other life-like images. It was trained for about 224,000 steps with cosine restart and a final detail run for last 10% with training biased towards higher ranks that significantly increased realism on a smaller, pruned dataset of very aesthetic images.

You do need to fight the anime anatomy with strong negative weights as a result of me not wanting to destroy/forget this knowledge, but I have a negative prompt used in all these samples which have proven reliable so far to avoid breaking down the realism with the bleed-in of big eyes, unrealistic head shape etc when you include anime characters and heavily booru-style prompts.

For numerous reasons I didn't purse any magic tags for different kind of photography, grain levels, camera characteristics etc so the aesthetic is pretty much what you get. If you want really fine, granular control over this then I would suggest other realistic finetunes or perhaps use/create LoRAs to get this control. I just liked this aesthetic, originally made the model for personal use and want to share it because why not!

The images supplied are all one-shot outputs, no facedetailer or hi-res fix.