It works very well.TLDR;<ul><li>kiryuuin ragyou, white hair, rainbow hair, multicolored hair,</li><li>OMIT the "kill la kill" tag</li><li>use the tags in order</li><li>The training data was regenerated and reformed</li><li>UPON REQUEST I will share information on the LOHA training Kohya configuration.</li></ul>OMIT the "kill la kill" tag unless you want to see her in the standard armored garb, which can't really be trained out without damaging all of the characters who wear it.Be sure to use the tags in order, as they have an activation pattern. They are fairly compartmentalized due to the training patterns and folder organizations used with LOHA, and the image size variations with bucketing make sure the system can regulate down to smaller images or even do some (fairly untested) inpainting. After the base tags, the majority of the sub tags were tagged using <a target="_blank" rel="ugc" href="https://huggingface.co/SmilingWolf/wd-swinv2-tagger-v3">wd-swinv2-tagger-v3</a> and shuffled for this particular version. The shuffle assists with interesting output, but can still cause occasional concept bleed. A more intricate tagging would also include <a rel="ugc" href="https://github.com/hukenovs/hagrid">hagrid</a>, <a rel="ugc" href="https://github.com/salesforce/BLIP">blip</a>, and <a rel="ugc" href="https://github.com/openai/CLIP">clip</a>, but this does not have that level of depth.This is both a benchmark test of my new training method, as well as an integration of a subset of poses I dub "mannequin" normalization. This concept has evolved from an earlier iteration as one I developed and only partially implemented in my original Obsidian model for 1.5. The pose system has evolved substantially and is the secret to a lot of consistency with this, so it will be used to train more models in the future.The training data was regenerated and reformed using multiple versions and pose interpolation. The original image data is based on images sourced from the internet (Gelbooru, Danbooru, R34xxx), and thus falls within the creative commons realm of copyright. That being said, you cannot sell created images from this subset, nor can you sell the image model as the rights are now mine mixed with the original creators.If contacted by the show's creator or the representative media corporation I will remove this model.UPON REQUEST I will share the why's to the LOHA configuration. It's a bit time consuming to train, but fully worth it. I will not share the training data, but I will share the folder pattern and importance of specific elements with the LOHA training method. The method is substantially better than any other method I've used, fits with various styles, does a substantially solid job at imposing it's own style, introducing new compounding character elements, and segregating various elements into their own compartments of access and importance. Any suggested improvements on this would be greatly appreciated, as it would save me time and testing. Keep in mind, while training, it looks like a garbled mess before it begins to form into something cohesive. Expect it to potentially look like garbage if you don't give it the correct sample prompts for quite a while.

Files with SHA256: f6052c5d364c5e10a6414852640e5a8da3686ee7bd3edf74289f435c21b46184