V1.0_ILIL models can generate her without LoRA but will either have the modern art style or some mismatches in her dress. So this version is made specifically to address my Pet Peeves.By far IL LoRAs are the most accurate to the original artworks in both details and style. (Surprisingly capable of other styles as well)This version did a bit overtrain on her royal dress. Causing her other attires to require more specific prompts.V1.5I forgot to set the caption extension in Kohya. So the previous version is actually trained without any caption.This version should be less overtrained and easier to work with other styles/LoRA.While it is now required captions to be more specific. But also there are way less background artifacts so negative is no longer required.V1.0A character from Brighter Than the Dawning Blue (夜明け前より瑠璃色な) seriesFor non-Pony LoRA. I recommend this person's LoRA <a target="_blank" rel="ugc" href="https://civarchive.com/models/268071/sdxl-character-9character-set">https://civarchive.com/models/268071/sdxl-character-9character-set</a><h3 id="character-tag-resr9i291">Character Tag</h3><pre><code>Feena Fam Earthlight, green eyes</code></pre>(May get blurry eyes without the tag)<h3 id="costume-tags-jy6q2w2af">Costume tags</h3>(Can add color/accessories to enforce more specific parts)Dress<pre><code>Princess, Dress, Tiara, Brooch, Elbow Gloves, Cleavage</code></pre>Sundress<pre><code>Sundress, Straw Hat</code></pre>School Uniform<pre><code>School uniform, Pleated skirt, White thighhighs, Bow, Emblem</code></pre>Bikini<pre><code>White bikini, Straw hat, Sarong</code></pre><h3 id="negative-tags-(optional)-utad6yqrr">Negative Tags (Optional)</h3><pre><code>Glow, Sun</code></pre>(The majority of training data have kind of glowy outline or a Moon in the background. Thus it tends to sneak it's way into the result if you try to have night sky and something similar)

Feena Fam Earthlight - 夜明け前より瑠璃色な (PonyXL)

I forgot to set the caption extension. So the previous version is actually trained without any caption. Less overtrained and should be able to be easier to work with other styles/LoRA.The downside is it required more detailed captions.

Files with SHA256: 7097d57c473ada5b9b576b23203102e2df9a72c5e4254e9d12c08e49577b1193