This is a LoRA for Sana Futaba from Magia Record.

While I previously used Illustrious-0.1 as the base model, I have switched to Illustrious-1.0 for this release.

The training image is 1672x957, so please generate it at a larger size, such as 1536x1536 or larger. Generating at smaller sizes like 1024x1024 will degrade the image quality.

I trained it on the anime version, but for some reason, the face doesn't quite look right no matter how I adjust the settings. According to GPT, she has few distinctive facial features, though it's also possible the model is being overly influenced by existing Danbooru tags. Interestingly, when generating images using just the series title and Danbooru tags (without the LoRA), the reproduction quality is surprisingly high given the limited amount of source data.

With WAI-14 or WAI-17 base models, the face resembles the original but the art quality suffers; with NoobAI-EPS-1.1, the face is accurate, but it lacks the "Full HD" crispness of Illustrious-1.0, so it's not to my taste. Pony resulted in a heavy, painterly style and failed to capture her likeness at all—the reproduction was terrible. Anima also produced poor results, creating blocky, doodle-like images, especially around the face. (Note: When I tested this with Iroha, the art quality wasn't an issue.)

Perhaps related to the facial likeness issue, the outfit tends to "stick" strongly, making it difficult to alter the uniform's characteristics; while some outfits don't change much, using emphasis tags like ((maid uniform)) forces a change. A weight of 0.7–1.0 is recommended. I suggest 0.9–1.0 for the "Kuroe" outfit and 1.0 for the "Pavise Shield."

I am making the dataset publicly available, so please let me know if anyone manages to successfully reproduce her anime-style face.