@DVNM3ZXAR Overall, I think it's really good. it successfully brings to life the character I'm looking for.

If you’d like some “constructive” feedback (and I say this from the selfish perspective of someone who only has a very vague idea of how to create LoRas), perhaps there’s a way to control the fact that applying more force doesn’t affect the style you’re trying to preserve in a specific generation as much.

This, of course, isn’t something that happens exclusively in this LoRa. It’s actually quite common.

At least the solution I use to apply character LoRas while preserving a consistent style is to use a “two-step” strengh application. First, I apply the LoRa with low strengh at the initial steps of generation to create the character’s “initial structure” (hair shape, face shape, etc.), and then, in the Face Detailer step, I apply a much more aggressive strength. This way, I can more easily recreate the character’s facial features without significantly altering the style.

Another thing that was particularly helpful was adding several hair tags to find that signature style for the character, using all of these tags: “purple hair, purple eyes, low twintails, short twintails, twin braids, hair flaps, hair over eyes, eyes visible through hair, double-parted bangs, crossed bangs, wispy bangs” to describe the hair. Otherwise, it would generate something similar to the hairstyle she wears with the maid outfit.

Here are a few examples: [⚠️⚠️🔞NSFW WARNING🔞⚠️⚠️]

This is the “Main” result, where I applied a strength of (0.20/0.60) (low-noise/face detailer) and used the hair tags I mentioned above.

https://civitai.red/images/144533487

Next, here are two more examples where LoRa is completely disabled, so you can see that while the model manages to partially identify the “kominami asumi” tag, it only partially recreates the character’s features.

One without hair tags

https://civitai.red/images/144534590

One with hair tags
https://civitai.red/images/144534589

Finally, here’s what happens if I use a LoRa strength greater than 0.65

One without hair tags Strength:(0.70/0.15) (low-noise/face detailer)

https://civitai.red/images/144534592

One with hair tags Strength:(0.70/0.15) (low-noise/face detailer)
https://civitai.red/images/144534591

But as I mentioned earlier, overall it’s pretty good. The details I’m getting hung up on are actually details that can be easily controlled by whoever uses the LORA, and the entire responsibility doesn’t have to fall on the creator who trained it, so I really appreciate this work. I’ve been waiting a while for someone to train this character.