my fourth attempt to make a LoRA based on Cyclops from X-Men, a rework of the SD1.5 lora
trained on 1024x1024 images.
Apologies for the large number of keywords, I was experimenting with seeing if I could separate parts of his uniform, and whilst that was a success it, a side effect also means you may have to use a majority of them.
I might try a figure out a way to have one keyword for the full uniform and the partial keywords for bits and pieces (maybe?)
Strength: I had it for 0.8-1.0 for the sample images
Trigger word: xmen_cyclops (including the underscore)
Other keywords: visor, blue bodysuit, yellow chest harness, yellow belt, yellow gloves, yellow briefs, yellow boots,
side notes:
(SD1.5 V1) blue bodysuit is a keyword, but it will almost always certainly generate the bodysuit if you have any of the other triggers even if you don't include blue bodysuit (after all most of the images in the dataset has him in the costume).
(Pony V2) has more images of Cyclops without his bodysuit so having, or not having blue bodysuit will have more of an impact
if you want him without the bodysuit but with the harness;, try nude, and yellow chest harness, as a keyword.
(SD1.5) You may have to place emphasis on nude i.e (nude :1.3) as when I tested without the emphasis it was a 25/75 shot of it working or just giving me a really skintight bodysuit.
(Pony 2.0) probably don't have to put the emphasis anymore as the v.2 dataset has more images of Cyclops without the bodysuit.
See the meta data on the sample images as the sample images go from fully dressed to buck naked.
the visor around the ears can be a pain in the butt, and you may need to fix it in inpainting (ramp up the strength)
(v1 - SD 1.5) Also uploaded to Tensor.art at Cyclops - X-men - 1.0 | Stable Diffusion LoRA | Tensor.Art
(v2 - Pony SDXL) Also Uploaded to Tensor.art at https://tensor.art/u/738805310864492090