Viridian's AI Experiments

[Viridian]

Here's a bit of a look at a recent new tool for NovelAI, which I combine with Stable Diffusion for better outputs. Recently, they implemented a "vibe transfer" tool. It allows you to upload an image, but use that as a reference for the "feeling" or tone of the image. Normally, you would use image references in "image 2 image", where you use one as a prompt to create another similar image. However, this meant that it was restricted in its outputs: it's harder to turn a red t-shirt into a blue bikini, for example.

With Vibe Transfer, I can use a reference image like this one.

vibe1.jpg

I can adjust how much I want it to stay close to the reference, but it will otherwise ignore the visual content and follow my written prompt, but try to replicate the style and intent.

I then input some prompts that will get me my Quicky Sanders character. Normally the output is quite unpredictable, but the vibe transfer narrows down the kind of thing I want. It captures the shouting, the intensity, the colour scheme and the dynamic hair. But let's say I wanted to change the entire outfit and character, and the setting to a sand quarry. After a few attempts I get this:

vibe2.png

NovelAI is, of course, an anime based model. To get it to my photographic style, I run it through EpicRealism on Stable Diffusion.

vibe3.jpg

I could, of course, have generate the image straight from NAI or SD without the Vibe Transfer step. However, it really helps with improving consistency in the style of output you get, which is always the hard thing with AI.

[Viridian]

Updates

[cerberus]

Good to see you back again and a nice set of updates as well. Excellent looking mud, very realistic and the girls as well.

[65sinking]

These are really great!

[sixgunzloaded]

Great expressions on all these. I particularly like the three red-heads, especially Runaway v2.

Going back to something Cerberus said a while back -

I suppose you could train a LORA for a specific face and use that in the prompt to get consistency

Is it not possible to just prompt "consistent face" and have it obey that? At least on an animation anyway?

[Viridian]

Is it not possible to just prompt "consistent face" and have it obey that? At least on an animation anyway?

Images are generated individually, so there's no such thing as a "consistent face" as it does everything one at a time. This isn't necessary though, as apps now have Face Reference functions, where you can upload an image as a facial reference and it will replicate that face.

I use the first image as the facial reference, and remade the other four with that facial reference.

[Viridian]

And face-swapping some other creations.