Illustrious version of this LoRA here: <a target="_blank" rel="ugc" href="https://civarchive.com/models/1604289">https://civarchive.com/models/1604289</a>First Lora I've ever trained. Well, technically second, since this is v2 after all. v1's biggest problem was that realistic images generated would all have the actress's face lol. Ended up deciding not to publish it since I want to make NSFW (but not that NSFW lol) images with it after all. v2 fixed the issue by cropping the actress's faces from the images used in the training data and adding "head out of frame" prompt lolololWould greatly appreciate it if you could read this before posting! <a target="_blank" rel="ugc" href="https://civarchive.com/articles/7060/look-both-ways-before-you-crosspost">https://civarchive.com/articles/7060/look-both-ways-before-you-crosspost</a>Main prompts:<pre><code>sazerpisces, armor, black leotard, silver bodysuit, belt, gauntlets, black gloves</code></pre>Add for anime images:<pre><code>source_anime</code></pre>Add if doing "full body" or "feet out of frame":<pre><code>boots</code></pre>For helmet (put "visor" in positive or negative prompt depending on whether you want face to be hidden or visible respectively; avoid using hair with helmet, and helps to add hair with high enough weightage to negative prompt to prevent hair from appearing behind the helmet):<pre><code>helmet, visor</code></pre>If you don't want helmet to be worn but rather held, then (though be warned that there is only a single image of this in the training data so your mileage will vary very widely):<pre><code>helmet removed, holding helmet</code></pre>For gun (not that accurate; maybe I should have gave it a distinctive prompt instead of a generic one):<pre><code>gun, holding gun, holding weapon</code></pre>Armor is removable by taking the following prompts out of positive prompts and adding them into negative prompts, but results are inconsistent due to only having one low-quality image of such in the training data:<pre><code>armor, gauntlets</code></pre>Note to self on potential improvements: 1. Add "from above" to training data; I suspect that's why discoloration of the costume occurs when "from above" is used as prompt 2. Instead of "helmet" for the images with helmets worn and "helmet removed, holding helmet" for helmets removed, add "wearing helmet" to the former and use "helmet" for both

Files with SHA256: 8887854b903c356ccb3477dbaef5d1ff3f3237674704eb5632a29faefecea4e0