Socalista BORDELLO Creator: Socalista
After quite a bit of data gathering, captioning, merging, training, remerging, retraining, and fine tuning, here's the first iteration of my Bordello checkpoint. This is a NSFW checkpoint, intended to make images of beautiful women in erotic situations. It is moderately stylized and will trend towards realistic scenes, people, and natural lighting. The goal of this checkpoint is to move further away from the dreaded "1girl" output. It is not focused on hardcore scenes, though will quite readily do them if you either use an appropriate LoRA, merge this checkpoint with a hardcore checkpoint, or use it as a refiner for a hardcore checkpoint to boost its realism.
Personal setting preferences:
Sampler: DPM++ 3M SDE Exponential, DPM++ 2M SDE Heun Karras, or DPM++ 2S a Karras OR LCM Lora with DPM++ SDE Karras or DPM++ 2S a Karras
Steps: 30-50
CFG: 3-6 - Lower CFG will do better skin texture/realism, higher CFG will do better ethnicity/prompt adherence though you'll get a little bit of that SDXL shine back with higher CFG
ADetailer - 0.15-0.35. Many prompts will not need ADetailer, especially if not zoomed out, but I just leave it on regardless
A few things it has been trained on that you can use in positive or negative prompt
round face, oval face, rectangular face, square face, oblong face, and heart shaped face - all have an influence on the output (and "heart shaped face" might give cute heart decorations or necklace charms. Just toss "heart" in negative prompt to resolve this)
thick eyebrows, thin eyebrows, angled eyebrows, arched eyebrows, natural eyebrows, curved eyebrows
wide chin, round chin, tall chin, small chin
wide jaw, round jaw - weaker results here, but can still get some shifts
prominent nose, small nose, round nose, upturned nose - weaker results here, a goal of future training
dark lighting, natural lighting, studio lighting, warm lighting, cool lighting, bright lighting
leaning, crawling, kneeling, bending
perpendicular, legs together, legs closed, legs apart, lying on stomach, lying on back, lying on side, turned left, turned right, turned towards, turned away, looking towards, looking right, looking left, looking away - weaker results, this is a goal of future training
This checkpoint works with a wide range of samplers, steps, CFG, and LoRAs.
NSFW details: This checkpoint might give partial or full nudity unprompted. This can easily be resolved by putting "Nude naked NSFW" in the negative prompt. It will very likely still give skimpy clothing, however. As with all SDXL checkpoints, to resolve the "nipples through clothing" issue, put "nipples" in the negative prompt. However, I have had some success with putting "transparent" in the positive prompt relating to clothing, and letting the nipples generate - there's a good chance they'll layer under the clothes instead of poking through. You might get random bra straps, if you do, just put "bra" in the negative prompt (many SDXL checkpoints have this issue).
Method: I took my Socarealism checkpoint, added in my own training data, and merged a few LoRAs. Initial attempts to do stronger stylization resulted in shinier/3d output and the return of the dreaded 1girl. Reworking it with more of a focus on my own training data plus lighting styles gave the results you see here. I splashed in a bit of the following LoRAs and Checkpoints in addition to what is in my Socarealism checkpoint: Low-Key Lighting Style, Rembrandt Lighting Style, RMSDXL Darkness Cinema, 35mm Film Photography, Trifecta XL, and Hephaistos NextGen Tensor Edition. None of these were merged in at high levels, so as to maintain the option for folks to use the LoRAs with this checkpoint without over-burning the image.
Current limitations: Trends towards Caucasian women (due to the initial selection of data I gathered - a well known site featuring Eastern European erotic models). The teeth could be a little better, I think this is a legacy of Trifecta, prompting for "open mouth" makes the model think I'm asking for a blowjob. Earlier iterations had this issue much more noticeably, but mixing in more Hephaistos and Nightvision stabilized it some. Also, due to the training / stylization, it is slightly less good at making celebrities than my Realism model - make sure to use a LoRA + Embedding (or appropriate keywords) and the issue will resolve itself.