DavesGaming

My ongoing Chuck and Sarah playthrough has been fun, and has required a bit of learning and experimenting with the AI art. Mainly because of my choice to recreate two favorite characters from a TV show. This led to me creating character models, mostly from photos of the actors. Because that setting was modern, I also created medieval fantasy versions of the characters in the gear I wanted them wearing (with Co-Pilot) that was mixed into their model builds. So in a nutshell, the models “know” Chuck’s signature light blue tunic and Sarah’s silver plate armor. This was a fairly involved process, but at least for the character portraits themselves, it has worked very well.
OpenArt remains a more limited generative AI I believe, but at least in that context of portraits it can do some very nice work. Photo-realistic renders of Chuck and Sarah as fantasy role-playing heroes.

I mentioned back when I did my Psyche and Diomedes run how special those characters had always been to me. They’ve been with me my entire adult life, over 45 years now. And in the odd way that makes sense to gamers (and maybe writers?) I know and love them.
That run was illustrated entirely by Co-Pilot. I got some good results, but there is a little problem of shift. Because Co-Pilot does not use models, each render is unique and the character is drawn from the input verbal description. I got some results I was quite happy with for Psyche. She was described as “Female Human, Age 20, Russian, short, beautiful, long brown hair” (sometimes “in a pony tail”). Dressed in either “purple robes with gold trim” or “sleeveless purple tunic” depending on how she was geared up.

These were the two portraits I used through most of the game, so we’ll say “definitive Psyche”.
But in many renders she looked wildly different.

Not an ugly picture! And correct in the written details. (I’d say she looks more Scottish than Russian? Ethnicity is a weak way of describing traits) But obviously very different from my view of Psyche.

So what if I put together a half dozen of my favorite Psyche renders and use them to build a model for OpenArt? What would a model built entirely from AI renders look like?

This is striking. Suddenly I have a much more consistent character. Its not quite as solid as my Chuck and Sarah models, I can think of two things that likely contribute. First, six training pictures is a pretty small sample, not a lot to draw on. Second, they aren’t completely consistent between each other. At least not like multiple photos of the same actor are. But seriously, this is pretty good.

A weakness of OpenArt emerges when we start putting these portraits into scenes. It is much more of a fight than similar scenes with Co-Pilot.
But Psyche *has to be* shown with her husband Diomedes. They have been an inseparable pair for their entire existence.

You can see the difference in these characters, Diomedes is simply verbally described like he would have been in Co-Pilot. While Psyche is a model. Obviously we see a lot more consistent features for Psyche between these renders than we do for Diomedes.

This was a fun experiment. At this point I have no particular goal in mind for this work. No future playthrough planned. But I’ll likely revisit these characters again as my abilities with the generative AIs develop.