A merge designed for generating good results from NSFW concept LoRAs. From my observations, concept LoRAs improve with training data from multiple mediums. So the goal of this mix is to blend mediums like anime, artistic, and realistic, but ultimately lean towards realism for output.

I personally use this model to test out my XL LoRAs. I don't feel it's as good as the original yet, but I can't keep posting my LoRAs without sharing my base for you all to test the prompts with 😜.

Generating

I found the best results were starting with 864x1152, clip skip 1, cfg 7, DPM3++ SDE Karras. Then img2img: resize by 1.5x, denoise 0.3, Script: SD Upscale -> R-ESRGAN 4x+ @ scale factor 2. (this eats up 24gb of vram for me).

No refiner is needed.

Prompts

Suggested general prompt:

detailed background, masterpiece, best quality, high quality, highres, absurdres, award-winning, professional, highly detailed

Negative prompt: FastNegativeV2

Suggested realistic prompt:

8k, highly detailed, ultra realistic

Negative prompt: cartoon, 3d render

Resources

Overall, this model is currently a merge of: Spectrum, Dynavision, Juggernaut, Explicit Freedom, Realistic Freedom, Gleipnir, All-in-One NSWF LoRA.

Caveats

Specifying any kind of breast size will likely generate nudity.

Specifying a hair color or facial features will make the results more human (though this may be what you want)

It has a slight watermark bias due to the merging of some NSFW LoRAs. If you're using LoRAs that tend to generate them, you can add "watermark" and "signature" to the negative prompt.