Explore your Digital Lust
Transforms most checkpoint into NSFW models.
Here is a experimental LORA for your nsfw needs. This model was created by extracting loras from the following list of model checkpoints and then merging them together.
I've tested it against a range of model, as you can see in the showcase. File is quite big, I will experiment with resizing later on.
Do not hesitate to share your creations, I'm curious about what people can create with the model! :D
Hints
I've been using DPM++ 3M SDE Exponential sampler with typically around 40 steps to good result. May work with other samplers. Use somewhere between 0.8 and 1.0 weight for best results.
Works particularly well with my other set of models (see suggested resources).
As this is built from model lora extracts, trigger words are very varied. Let your imagination run wild and test the limits.
Versions
V1.0
Extracted LORAs from the follow models and then merged them (in order of weights)
Description
See description
FAQ
Comments (6)
its pony inside ...
tongue and big eye ^^
What do you mean. The model does have anime content included which explains the big eyes and what not, but unless I'm mistaken, no pony was included.
Nice Lora, teaches many SFW checkpoints some NSFW and porn content :) I like it very much
Most if the time the face is not changed to much but I would hope that the look of the face made by the checkpoint is not altered to much
I tried to extract a useful LoRA out of my own model before and didn't come up with anything all that great. You've got to tell me how you managed to pull it off and merged it into something good!
The isn't much method to the madness (+ that was done a while ago). Mostly try something, test, try again. Wish I had more insights :(
This is really great for models like Rumble or Iniverse, which are good with X but lack XXX knowledge.
For those types of "softcore" models, you only need around 0.4-0.6 strength to turn them hardcore.
For general-purpose photorealistic models, with minimal X and no XXX , use 0.6-0.8 or so.
Around 0.8-0.9, it tends to affect the style a lot, and tends to make everything a bit too porny (fake tits, heavy makeup, high fat, watermarks, plasticky, oversaturated, etc). In that case, just lower the strength. It doesn't start to really fail until you go under 0.3 or so.
For saturation/contrast issues, use Vectorscope, or just manually edit before upscaling.
Like with all photorealism, you generally want to use a DPM sampler with Karras scheduler. Despite all the advancement in recent years, it seems like Karras still produces the best skin textures.
It's also useful to experiment with only using this lora for the hires pass, and not the base image, or using different strengths for base and hires.
The shotgun approach of using fast samplers for the initial image, then using higher quality samplers for i2i/hires is the most time-efficient, assuming you're generating locally (which everyone should be doing, because generation services might cease to exist tomorrow. Also, save everything locally and never rely on any site existing tomorrow) and not paying per generation.