Stable Diffusion (referred to as SD) is totally free with a lot of “getting started” tutorial support such as stable-diffusion-art.com where I started a few weeks ago.
For a good experience rendering, although not necessary, you might wish to run it on your own graphics card with good GPU power. I’m running it on my gaming rig with a 2018 era Asus nVidia 1080p with lots of memory and flops.
SD doesn’t do porn, in fact it has been disabled by the developers. However a group of folks are working on Unstable Diffusion to overcome this issue. They have a discord channel (that may now be under review).
SD uses checkpoint files, which I think of as a database of neural training. There are a lot of these checkpoint files (see tutorial mentioned above for a few of them).
SD works by constructing an image based on guidance from a prompt, such as “A photograph of a Tiger butterfly on a rose” with occasional help of a negative prompt such as “animal-” to exclude the animal tiger. The problem is, if the checkpoint file has only been trained to recognize flowers, it can’t draw a butterfly because that term is not in its vocabulary.
However, you can merge checkpoint files, and get image knowledge, or influence the image growth in another direction more to your interest. SD uses a seed value to create images, so if you don’t like the result, run it again with a different seed. Sometimes I run batches of 200 images to find a few I like.
There are many more image construction influencing factors, some as trivial as the physical final image size. Unlike DAZ with a bazillion precise dials, it’s impossible to get the exact image you are trying to achieve, and constructing the exact prompt language to get you even close is hard.Most of it is trial and error, the complete opposite of DAZ or Poser, so it takes a different strategy.
Finally, SD 1.5 has a lot of trouble growing hands, so 90% of the human form images that have hands that are visible in the image are just wrong. In all, as a newbie, it took me about 7 hours to craft the language for the “A very hot hot shower” using my personal checkpoint blended file, and I generated, conservatively, at least four hundred 600x900 images. But in the end, the renders can be quite amazing.
Financial opportunities: Just guessing, but seems in the future an entrepreneur might sell a custom trained checkpoint file for a specific genre of image creation.
If you look closely at “A very hot hot shower”, she has a deformed arm that is obscured by the shower spray LOL. Here is the spec:
Prompt: an uncensored hyperrealistic photo of a ((beautiful)) (((standing)) wet nude (((young))) adult female taking a fine mist steamy shower, soft focus, cannon 5D, ((blue)) shower with silver fixtures, long auburn hair, (((arms at sides))), stunningly beautiful highly detailed (wet) face of REDACTED, erect:10.0 wet nipples,
Negative prompt: (((ugly))), disfigured, deformed, cropped, long neck, deformed hands, cross eyed, monochromatic, large:10 breasts,
Steps: 30, Sampler: Euler, CFG scale: 7, Seed: 709845088, Size: 600x900, Model hash: b5f53f94,