@kenney what I hate the most is that these tools can be used to create decent graphics if one puts in the work and the individual graphics show that .
however the people making these asset packs don't put in effort to make the whole graphics packs good quality .
like , the moral background of how the models have been trained aside for a moment , I think that they could be used to make a good image pack . could I do it , at least to prove a point ? no ! I both don't have the skills required to get image transformers to output good looking images nor the knowledge to compose a good asset pack . these skills do impact the quality of the work while taking effort to develop and I think that should be acknowledged .
if someone , who not having an ability to create art themselves , makes a asset pack which is consistent , isn't just topic vomit and most importantly is good and usable in the context of creating a game , I don't see an issue with them selling it , as long as they disclose they used a text to image model .
that said the way the current generation of image transformers has been trained is not that ethical , though admittedly I am unsure on how ethical it could be , at least under the current socioeconomic system
I hate how all of this very interesting and powerful technology is taken by people who don't fully grasp its limitations to then exploit for financial gain while over blowing its capabilities
@kenney The biggest giveaway is however can be known from how the AI works in simplified terms. It uses statistical probability to calculate what pixel comes after the next within a certain context window.
Image generation AI is just a very spicy Wave Collapse Function Algorithm, and knowing it makes sense out of all the mistakes it makes (weird fingers, background breakage, messed up patterns, etc.).
@PixelPerfectEngine @kenney I'm not sure if that is correct, although I'm not an AI export either… My understanding is that image generation AI works on the image in a global manner, removing noise everywhere at the same time in each iteration, so there's no "context window"…?
Otherwise wouldn't we be seeing errors like "infinitely long fingers" rather than just messed up structure (which can also be explained by the AI just ending up in some sort of search space local maximum)