v2 update on 20 May 2025:In v1 of this Pony LoRA, the cat ears portion of the helmet tends to keep appearing on the heads of characters in images generated even when you don't intend for the characters to wear helmets. I made changes to the training data by changing the images without helmet from "head out of frame" to "faceless, censored identity"For some inexplicable reason though, the quality of this LoRA took a sharp dive. Most of the time you get low-quality images generated out of it. Nowadays for training LoRAs I just take the latest epoch and run with it, but for this one I actually downloaded multiple epochs and tested them to figure out which one gives the best output, which, unfortunately, is still not really goodAlso it doesn't really fix the cat ears issue. Mitigated it, definitely, but I wouldn't say it definitely fixed it lol. As again, "wearing helmet" in negative prompt at high weight would helpWhat I'm trying to say is... Use the Illustrious version of this LoRA. It gave me much less problems overall. Sure, Illustrious isn't that good at costume accuracy, but I would rather take costume inaccuracy over generating, cancel generating, generating, cancel generating, generating, cancel generating, rinse and repeat lol<hr />Made my own Sazer Visuel Lora because the previous one I uploaded (which was not made by me) could not depict the lower half of the costume accurately lol, hence why most of the showcase images are focused on said lower half hahahaMain prompts:<pre><code>sazervisuel, orange bodysuit, brown leotard, belt, armor, gauntlets, black gloves</code></pre>Add if doing "full body" or "feet out of frame" (training data set had "boots" as prompt, but from my own testing, "knee boots" seem to get you a more accurate depiction of her boots including her silver knee guards):<pre><code>boots</code></pre>For helmet with face covered:<pre><code>helmet, visor, wearing helmet, covered face, covered mouth, covered eyes</code></pre>For helmet with face visible:<pre><code>helmet, wearing helmet, [eye color prompts], [facial expression prompts]</code></pre>For helmet to be held and not worn (I forgot to actually test this but it should work since there was an example of this in the training data set 🤣):<pre><code>helmet, visor, helmet removed, holding helmet</code></pre>For her claw weapons (which AI seems to struggle to understand what claws are despite images with her wielding them being included in training data set, so your mileage may vary):<pre><code>claw \(weapon\)</code></pre>Notes:<ol><li>When generating without helmet, the animal ears portion of the helmet may appear. Including "helmet, animal ears" in negative prompt mostly helps when this happens, but seems like not always</li></ol>

Files with SHA256: f1e807ac2f18131d9453bcd42eaaf98d6e5be1def52b3d2240ce56429d47b662