Hello, this is one of my first LoRAs.
Most of the dataset I had hand-drawn myself. I had to grab some references from web, but then modified then. I added some weird stuff on my drawings. Hence it may render with some weirdo features.
It can render him for some film-like capture-esque, or for NSFW.
Some of the dataset includes him with ears and short hair. On my drawings I paid special attention to well drawn male hands, fingers and fingernails.
For prompts, you can use your imagination. It may or may not work (again it was a small dataset optimized via Prodigy).
I used anime tags, for clip ones were very lacking in details.
May work best with semi-realistic, 3d, cartoonish and similar models. I haven't tried with realistic ones. Most of the dataset includes pictures with 3d-rendering feels. Which means it may change drastically results depending on the model.
Despite having him wearing his hockey mask on all pictures, some models insists in render his nose unmasked.
Usual prompts:
quality tokens: best quality, masterpiece, etc;
face covered (by mask)
If it insists in not drawing the mask, you can add "male hockeymask".
For negatives:
quality tokens
you can add "unsymmetric mask", "uncovered face", "uncovered nose"
No need to add negs like "watermark, signature, etc", dataset has no signature, not even my own (since many are handmade digi-illustration made on gimp with wacom pen). Unless the model you use is filled with those things. Bear in mind though, that "highres-fix" sometimes interprets elements from the picture as text AND ADDS logos, signatures, names... negatives won't prevent them. You will have to remove them with your own means.
Despite having some dataset with him holding weapons, as the classic machete, not all models, or not all the time it delivers them, instead they take him as a hockey player. You may want to leave it like that, or modify afterwards.
Pay attention to some samplers used in highres-fix as some of them, when set too high may remove his mask and turns him out with a normal face (randomly picked by model of choice).
Since I cannot link , useful (if you find too many distortions) embeddings are:
Useful Quality Embeddings
badv5
deformityv6
aid (in many versions)
whatever that user posts, they're awesome
Unspeakable Horrors (Negative prompt)
that once related to composition and its v4 should be enough
Strength can be set somewhere between 0.6 - 1.0 (And 0.9 or 1.0 on Adetailer).
On ********* it seems you won't need Adetailer that much - depends on the model of choice though.
You can read full description on Civitai at:
https://civitai.com/models/190954/j-son?modelVersionId=214463