The state of visual generative AI is quite laughable at present.
I would never use it for anything of substance, on principle, seeing as it is based on stolen data, or at very least content obtained through legalese and broken promises. Nevertheless, just for giggles I tried to get Dall-E to produce something to a very specific brief. Despite giving it 4 further refinements (the daily limit, apparently) it was still inadequate.
=> More informations about this toot | More toots from realcainmosni@mastodon.me.uk
The thing is the prompt refinements were not for aesthetics, they were to correct blatant flaws in the output, and every "fix" briught with it other errors which got progressively worse. The thing gets about 80% of the way there, but gets the other 20% stunningly wrong: two tables when one was requested; decorations when they were explicitly deprecated; six-toed feet; an ash tray on the floor not the table; a table with a doily frill materialising around it but a polished top... it's dreadful.
=> More informations about this toot | More toots from realcainmosni@mastodon.me.uk
text/geminiThis content has been proxied by September (3851b).