X2: Only works in my XLP X2 model.

Use 8-12 step turbo karras workflow. Inpaint the head/shoulders area at 1 strength to correct inaccuracy. Most details will need to be inpainted since it's a low rank lora in an inaccurate detail turbo model.

832x1216, 768x1344 (full body), 896x1152, 1152x896 (lying horizontal shots)

Just a test to see how far likeness and fitting could go in it using a suboptimal dataset. Training for X2 works good starting with high batch size AdamW then switch to low batch ADAfactor when it's well fit to reduce adverse noise and hone features. Dataset is bad quality but gigapixel restored, piercings removed, and de-wordmarked. There isn't even a relatively SFW image of her anywhere so this is obviously a very cleavage focused model because of that and a lot of the likeness is tied to those tags. Strongly responds to negative prompts and they need to be used strategically. The string I used seems to not interfere.

Dataset keywords: long hair, closed lips smirk, mega-breasts pushup, areola_peek, head turned to side, looking away, upper body focus, self grab, overhead angle, low angle, slightly lowered angle, fisheye lens, no_nude, tease, mole under lip, intimidating look, hanging breasts, hair_censor, harshlighting, rimlighting,

Outfits: semi-open_shirt, shirt unbuttoned from bottom, underboob, rainbow knit, purple strappy lingerie, knit low-cut bikini, cutout, jacket, visible nipples sheer, corset, large hole cutout bodysuit, thigh garter, mesh, highleg panties, sheer top

Strong keywords:

outdoors - use this in the negative the crappy cellphone nature backgrounds introduced all of the adverse noise

supported_breasts - with bra or supporting clothing

unsupported_breasts - natural hang and sag, use either one of these and there is no need for size prompts they are accurately sized and proportioned.

original_look - those old photo filter shots she started with where she looks much older

photo_filter - supports the above

cold_tone - a cold washed photo filter

warm_tone - all of the sunlight an yellow indoor shots

SD 1.5:

LCM method which always saves bad Loras. Turn it up past 0.8+ and use the basic prompts and find a good seed on 512x512 or 480x560 or similar sizes. Then lock and use Hires fix up to 2x or as big as it needs to be to get her face in accurately.

Adetailer can be used in normal non-LCM ways, Both face and person.

Other than that you get underfitted model images because this is scuffed. I have a new way to train exact likeness in way better and I have better buckets and consistent dataset images to try on V2.

Activator: LKvanski, RAW photo, Ukrainian model, cleavage

Dataset keywords: bikini, strappy outfit, seminude, cleavage, full body, sheer top, ass