Digital Styles

So, recently I've been playing around with different mixing art styles, with a focus on digital, semi-realistic painting styles, and I came up with four that I liked.

UPDATE!

I was running into some issues with v1-3 (most likely from slight overtraining), so I changed stuff around and remade them. If you haven't had any issues feel free to keep using the previous versions, but I've replaced the files with newer versions now.

These new versions mostly maintain the same styles, with only subtle differences.

v4 wasn't changed, as it was a merge of the other three at low enough weights to avoid the same issues.

Added v5 which is another low weight merge based in these newer versions, and has it's own style.

(I apologize for the pale, ghostly girls)

The styles in these samples don't look so different, I know, but this was just based on a trimmed down version of a prompt I was using before running the samples.

More complex prompts show a better variety (You can see the other posted images).

Anyway, a brief breakdown, from left to right we have two samples for each style (portrait and torso).

v1 has sharper details, and a slightly yellower tone (especially with lighter skin) from what I've seen in my testing.

v2 is the opposite, it has softer shading and a lighter, slightly hazy quality to it.

v3 is similar to v2 in that it's a bit lighter in tone, but is has somewhat sharper details and a slightly more stylized look.

v4 is most similar to v1 in terms of sharpness of details and overall realistic quality, but still has a lighter color tone.

v5 is close to v1 as well, but this one does have more shine to the shading.

(I changed v3 to 'stylized', v4 to 'smooth' and name v5 'shiny', though these names are only for CivitAI, the actual loras just have the version number)

All four models were trained on an alternating dataset with some minor overlap. I'd simply replace most of the images with a different style, train the next model, and repeat. I did this because I was going for a softer semi-realistic style, but wanted to try mixing other additional styles with it.

v4 and v5 are based on a (low weight) merge of the first three models, along with training with a smaller version of the dataset. Both still maintained their own styles so I kept them.

As for prompts, the main thing to note is to use "realistic" in the prompt for the semi-realistic style, without it you may get something more 2.5D anime/cartoon-ish, which still gives interesting results. (This is also dependent on the base model you use; since I use WAI it gives the more anime style.)

Adding "digital art" or "traditional media" to the prompt also helps improve the overall painting style.

All five models were trained using the Illustrious-XL v2.0 checkpoint.

All generations are made using WAI--illustrious-SDXL v14 and Fooocus, and upscaled to 2048x2048.

All of my models may be used for your own training and/or merges.

I also started using DeviantArt for posting my AI generations, style tests and even a look into data sets for certain models, https://www.deviantart.com/zabadugremlin