Game Art Feedback Wanted (RPGM) (StableDiffusion)

[ChonkyLove]

Hello my fellow deviants,

I'm making an RPG Maker game and wanted feedback on the character art. The game is focused on a lazy college grad who moved back home (original, right) and his landlady who is frustrated at his lack of direction in life and suspicious about his intentions.

The primary difference between my game and the many other similar plots is that I'm using StableDiffusion to generate my images, so I'm going to take advantage of the fact that I can generate many high quality images at once and stack lots of content into the game.

Before I start hammering my GPU, however, I wanted to ask around and see whether this lady looks good enough for a middle-aged, thick landlady.

Cheers, y'all.

**EDIT** These are not final images by any means, just the current parameters I've been playing with.

 

[levitz]

She definitely looks good generally, but there's some work to do. Fine tuning as you might say.
Overall: i like it and just keep in mind i would've liked it more if i didn't knew it's ai *wink

 

[ChonkyLove]

Overall: i like it and just keep in mind i would've liked it more if i didn't knew it's ai *wink

I feel this but I was considering being open about being able to regenerate the images for the game per people's preferences. For example, if someone played through it and was a fan but asked if I could make her skinny and more eastern Asian looking, I could tweak some stuff in my massive prompt and generate a new image set.

 

[osanaiko]

I think you've done quite well at keeping a fairly consistent visual look for the shape of her face and hair - that is one of the hardest things with the current generation of SD generators. the 1st and 3rd are especially close, the 2nd one is a bit different.

However the inter-image inconsistency issue (*) is showing by the way the background is changing on every image - she's standing in different kitchens (different dishes, shelves, faucet shape). You could potentially reuse a single background by snipping and stitching the different foreground female images into a single background and then redoing a img2img pass with a low CFG value to clean up the foreground/background borders.

The other big problem is the hands, they so often come out very mutated with SD1.5 or 2.0. There are various techniques for improving the hands gradually being discovered (e.g. depth maps and pose generator) but you'll have to keep working on it, such as doing repeated in-painting with hand depthmap.

At this stage it's still a lot of work, but I guess it is much less than the several years it would take learning to draw that well.

As we all get better at AI art prompting and workflows, and the tools improve, I hope AI generators get good enough to allow me to try to do the H-game of my dreams.

(*: whoa that's some serious alliteration)