Synthetic ??? SDXL

This is a LORA intended for mature audiences only. Do not abuse this LORA. Don't use it on real people, and do not use it for any illegal activities.

V2 UPDATE: Trained on real ??????? images instead of synthetic ones. I inpainted away superfluous body parts (and ????) so this seems to be coming out as intended now. This model was trained over 6 epochs (vs 9 in the prior) and seems to be happy around a LORA strength of 1, but you can still mess with it of course. I optimized for enhanced/photographic shots with lots of token terms, so you should be able to make up your own scenes.

Cum is a lot more accurate close up, but that's no surprise. Remember to up the number of steps if it starts looking distorted. It's freaking ?????, so you're gonna see a lot of ???????. Only seems to work well on women.

V1 UPDATE

This is an experiment. I took a ??????? LORA, added whipped cream, generated a bunch of images using a model that, for whatever reason, generated the most coverage, and trained another LORA from those images.

IMPORTANT

• You'll need at least 40 sample steps for this to work. No idea why, but if you start seeing weird colors and slime, just up the sample steps.

• If you're getting closed eyes, add "eyes closed" to your negative prompt. I tried to engineer this out of it, but it seems to crop up still on occasion.

• This is a ????? LORA, so add "?????, ????, ???????, ???????" as negative prompts if you're trying to do e.g. a ?????? or whatever else.

• You're going to need to scale the LORA back to at least 80% (maybe even down to 60% depending on if you're going for lighter ??? coverage) for close up shots, e.g. <lora:synthetic-cum-v1:0.8> - this model is gonna push you in the "close up" direction, so try varying the LORA strength based on how close or far away of a shot you're going for. The problem is the higher the LORA strength, the more it's going to ignore e.g. "Full Body Shot", and the weaker the ??? effect is over distance.

PROS

• No ????. It's been all cropped out. Stable diffusion can't handle the D imho.

• Lots of ???, can be scaled back for higher quality image generation.

• Good at close up shots.

CONS

• Over-trained on Asian girls. Add "Asian" as a negative prompt to offset this.

• Freaking ????? LORA - most images are coming out ?????, even with "?????, ????, ???????, ???????" as negative prompts.

• At 800mb it's a bit beefy.

• I used WD14 captioning on it, so pretty much anything will trigger the LORA. I need to do another version with natural language.

• Limited range of angles - the source dataset for the original LORA had mainly close ups, so the LORA struggles at distance especially.