Subject: That is, unfortunately, one of the limits of the AI thingies today.
Author:
Posted on: 2022-12-08 21:15:32 UTC

While they're steadly improving, they of course struggle with consistency. Once you have a character with, say, green eyes, the WaifuLabs AI doesn't understand that the character must keep it - it understoon that the image is "mostly right" and knows that there's some green there... but not if it is right.

I'm currently dabbling with a more sophisticated one (Stable DIffusion, which I tried with Waifu Diffusion, AnythingV3.0, and an anime-focused derivative of NovelAi as the models. A model is literally the "scheme" the AI uses to interpret your written prompts) and, despite being lightyears ahead of WaifuLabs it still shows that kind of behaviour.

Easy example from one of the random prompts I had given it: one of my agents, Nikki, swimming underwater, using the modified NovelAi model (AnythingV3.0 tends to get some details like the hands better, but also has a creepy preference for making her look exactly in her mid teens, and Waifu Diffusion is definitely the worst of three in term of resutls.). The prompt, of course, included the fact she has green eyes, but I put a very generic bikini as another prompt just to tell the AI that she was supposed to be in a swimsuit and not, say, in a dress as the AI doesn't know that people usually swim in swimwear! The unexpected result of the 5 images batch (maximum my GPU can handle in a single batch without running out of VRAM) I had told the AI to make?

promptone

As you can see, the AI has quite the preference for giving Nikki a green bikini, as the AI had no clear instructions for the bikini's color, but had other instructions in which green was a desiderable one and went for it. However, Nikki's favourite color is white, and she does tend to go for swimwear at least partially in that color. So in the following batch I specified that I wanted a white bikini.

prompttwo

And Nikki got her white bikini at last! Except, as you can see, when the AI thought to add a second color to some details it went straight for green again in three cases out of four (wth the fifth being a fully white bikini) The only case it used a different color (bottom left), it took the same shade of blue as the background water.

On top of that, as you can see there's still a lot of graphical "glitches". But it is still very interesting stuff to play with! Only thing is, as the models are "trained" by feeding them thousands of actual drawings to analyze, the jury is still out on whether the AI is truly "taking ispiration" or is simply assembling an image from thousands of stolen and edited bits. As such, I'm currently making no public use of my generated images - the two batches I posted here are literally the first time I shared my own Stable Diffusion-generated images online outside of private Discord chats.

Reply Return to messages