Rough draft of ideas involving AI image generation.
(Ignore this as I'm spitballing ideas)
There are a few things that define image generation and the final look images get.
1: art style
Art style (artist tags, art types) heavily define the look of images. A good example is something extreme that you want to generate like gigantic breasts (not human sized macromastia large, like excessively large) default tags just don't generate that. A fix to tard wrangle the AI is to mix art styles to force unique generations.
When using art style tags objects and locations will also change to fit the style of that artist. If an artist draws a bedroom the bedroom location generated will look like that bedroom. So good for spicing up locations, bad because there a bias given to the generation.
2: objects, clothes, things
Pretty easy to understand. Bikini implies beaches or pools. Building implies cities. If you use the tag of a known character using tags of the clothes they wear it will make that piece of clothing look like what they wear.
3: angles, image composition, poses
This has the least effect as poses don't really change an image dramatically. In fact the AI will "guess" the appearance of a character even if that character has only been seen from the front. The only issue with pose tags is that they will sometimes cause extra limbs or deformities.
The key to these tags is that they are used to make specific images that the AI has a hard time making. For example everyone loves mating press art but the AI has a hard time naturally making it. In some cases a "from below" tag had to be used even though there are no "from below" images of mating press in Danbooru. The reason being is that "from below" is helpful in angling the ass towards the camera.
I'll explain more about this issue later. #zemAI :back_from_gab:
Add comment