Stained Glass Immersive

v1.0b

This version is "b" because I don't consider it to be an actual improvement with respect to the previous LORA, but more of an adaptation to match my own particular taste. You may like one over the other, but this is closer to the style I like.

The new dataset was built so that character looks more like part of the stained glass, and to have a texture more close to actual stained glass. The downside is that I wasn't able to fully replicate the texture with this LORA, but its close to my training images.

As before, I trained 2 versions: one with WAI (12.0), and one with Illustrious (1.0). I like better the results of the Illustrious one (with Illustrious), but obviously WAI is more versatile. My ideal setup would be to generate a base image with WAI and then refine it with Illustrious (each with their own dedicated LORA).

Again there are two trigger words: stained_glass (to override the original concept) and stained_glass_immersive. Look at the documentation of v1.0 for further details,

v1.0

An evolution from my previous LORA. The objective of this LORA is to give a more stained-glass-like appearance, specially talking about the background behind the characters. As my other LORA, the main goal is to make the characters look like they are actually part of the stained glass --thus the "immersive" part (mainly the face, since most stained glass models out there --if not all-- tend to show plain faces without the LORA affecting them).

I decided to post this LORA as a new model instead of a new version because of two reasons:

1) The other LORA actually provides 6 styles in one LORA, while this one specializes just in improving one of them.

2) The dataset for this LORA was created using this one as help.

The dataset consisted of 44 images (1024x1024) as close as possible to my ultimate unique style goal (yet a way to go to bring into reality what I have in my head haha). Up-scaling my help your generation, specially in full-body images, but it is not required.

This time there are two trigger words: stained_glass (to override the original concept) and stained_glass_immersive. The weights of the LORA itself and the trigger words have worked fine at 1.0 during my testing.

I tried to bake the style of different artists in the dataset to avoid overfitting to any artist's specific style. This LORA thus works fairly well with artist's style's tags (but you may want to reduce the weight of the artists, since there is a trade of between my stained glass style and other artist's styles). The LORA works fairly well with other LORAs too, but be cautious when using too many concepts at generation.

DISCLAIMER: I think my favorite generation using this model is using JANKU as checkpoint, although training was done with WAI v12.0 and Illustrious v1.0 respectively.