<h3 id="v2.2">v2.2</h3>Pretty much v2.0 fine-tuned to increase emphasis on the lead structure. This time the LORA is more robust to good quality tags, but the more you use, the more "anime-drawn" the image will look (and less stained-glass-like). Again, use sketch as negative prompt for the better texture. For further info, read v2.0 below.IMPORTANT EDIT: Playing with <a target="_blank" rel="ugc" href="https://civarchive.com/models/1617606/stained-glass-mosaic">my previous LORA</a>, I found out that both that LORA and this one work particularly well together. This LORA should be kept with a weight of 1, while the other one works fine between 0.3 and 0.5 (I prefer 0.3). Increasing both CFG and steps helped me to get better results, so I recommend 7 and 35 respectively. In this <a rel="ugc" href="https://civarchive.com/posts/27583912">new post</a> you can find my results (using the same seeds as the <a rel="ugc" href="https://civarchive.com/posts/27512739">original post</a>).<h3 id="v2.0">v2.0</h3>So this version is 2.0 because it provides an improvement in texture and lightning with respect to 1.0b. 1.0b was fine for the lead structure, but the texture and lightning didn't look similar to glass. Version 2.0 aims to add stained glass (granular) texture and uneven reflects as if glass was hit by different sources of light on both sides (considering that stained glass is semi transparent). Although the lead structure got a bit affected, I think the result is good enough to share as a brand new version, hoping I can train a better one in the near future.Trigger words are the same as before. For any reason still unknown to me, I need to add sketch as negative prompt to get a clean version of my texture. Without the negative tag the result is not actually bad, but it doesn't really resemble my training data. Since the glass texture is a noisy, "bad" quality one, it's preferred to avoid good quality tags such as "masterpiece", "best quality", "amazing quality"... not that they suppress the LORA effect, but I find they try to "improve" (smooth) texture and "clean" (reduce lead lines) faces and hair.Increasing the number of steps may improve your results, since this is kind of a brand new concept that goes against most (if not all) styles baked in any given checkpoint you use. Unusual glass colors may sometimes bleed the skin of your characters; using "colored skin" as negative tag can reduce this. Look at the other versions' descriptions for further info.<h3 id="v1.0b">v1.0b</h3>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, <h3 id="v1.0">v1.0</h3>An evolution from<a target="_blank" rel="ugc" href="https://civarchive.com/models/1617606/stained-glass-mosaic"> my previous LORA</a>. 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 <a target="_blank" rel="ugc" href="https://civarchive.com/models/1688326?modelVersionId=1910813">this one</a> 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.

Stained Glass Immersive

Trained using WAI-NSFW-illustrious-SDXL v12.0 as base model.

Stained_Glass_Immersive-WAI.safetensors

v1.0-WAI

Files with SHA256: 0fff9d237e9191f9e85be3dfbc38522288d4dd1b430bb3aa96757d85f4f1cefe