This model can do anything ranging from any art style/photorealism/nsfw/porn/moderate text with ease without having any bias. It is the result of months of experiments.
You don't need any lora for porn or whatever.
The provided examples are "what you see is what you get" 24-30 steps with dpmpp3m sde/unipc.
Patreon link: extraltodeus. I don't lock anything behind it so do subscribe to keep it that way!
My goal is to create a balanced and poly-capable model.
Let's say you have the salaries of 30 people.
But among them there are 4-5 that makes from 50k to 2M per month: The average value is not going to be representative of most of the people within your group.
How can you determin then what does the majority makes?
Well lets say that for each salary you take the "distance" to each other point (so subtract and just keep it as a positive number). You keep the average.
Now for each point you have their average distance from each other.
Subtract to each of these the smallest of them. Now for sure the smallest distance is 0.
Divide them all by the biggest value. Now for sure the biggest is 1 and all the proportions are the same. 1 being the furthest to the others, 0 the closest.
Now for each do 1 minus the value. Uno reverse: 0 becomes the furthest, 1 the closest to the others.
What if you multiply them by themselves now? For example 0.5 squared is 0.25.
Power 4? 0.5 becomes 0.0625. Neat!
Now divide them all by their sum. For sure your sum is now 1.
Use each of these final scores and multiply the original values. Sum them: You obtain the consensus that has eliminated what was the furthest from the others.
Now imagine that a models layer is a simple grid of numbers, take the same layer for 30 models for example. Do the same computation for each individual value. That's how this model can write better and overall can be coherent througout any subject.
I don't clame Iris Lux to be perfect at all but that is as close as I could get.