Home Artists Posts Import Register

Content

So, I thought I'd poke around a bit more with the AI Art software to see what I could generate; a little bit of downtime fun inspired by the recent short stort writing.

I'm using Artsmart here.  I started with Duncan, the protagonist of the story, and wondered whether I could generate a coherent male and femal version of the same character.  It took a bit of work to get a suitable Duncan: the AI really went ripped-abs, romance-novel cover on him and initially refused to give him a shirt.  I used a range of keywords, like Scottish, warrior, wild landscape, etc.  At first I tried to get an imagine of him holding an axe, but the AI just found that too difficult.  I finally tamed it a bit by adding "in the art style of Angus McBride," and that seemed sufficient to get Duncan to finally put a shirt on.  You can see pre-shirt and post-shirt versions of Duncan here.

Then it was time for Aubriella [spoiler!], Duncan's transformed identity. No surprise, the AI struggled even more with it.  At first it went well.  I took the image of Duncan and fed it into the AI as an "inspiration" and that worked, but then I wanted to generate something closer to the long description in scene 1, where she stands in the chapel of Submission and looks at herself in the mirror.  The AI simply refused to: 1) give a full-portrait image, and 2) put her in a dress that went from neckline to ankle; it insisted on slits up to her waist.

Eventually, I realised that any keyword like "stocking" or "corset" was to blame; it can't manage the idea of a dress over shapewear and instead just forces the item to appear.

What ended up working best was 1) giving it the original image of Duncan, 2) removing the random seed generator and using the same seed, and 3) pruning back a lot of the language.  I copy and pasted the long description then pared it back and experimented with different words and eventually managed a handful of decent images.  I think they manage a certain coherence - male and female version could conceivably be related to each other - and the clothing and background hew close-ish to the prose.  

In any case it was fun playing around with it.  I'll probably try my hand at a bit more in the future, but for now I'll leave you with some of my attempts.  Which ones do you like best?  

Files

Comments

No comments found for this post.