V2 is now available. It should have fixed all the pain points from v1, making this version ready to move out of the prototype phase!The model should feel much more accurate and consistent now. You are unlikely to see stiff, overfitting faces. You can now put Ibuki Maya in all kinds of outfits, as long as you know how to describe them. See the showcase images for ideas!What I fixed in v2: I retrained the model using Prodigy with the following parameter: "weight_decay=0.001". This turns out to be the key of training a consistent Flux anime character LoRA!Original notes for v1.1:V1.1 is out - it is trained with much lower network dim (2!) to improve stability and consistency. I also trimmed the dataset to remove some noisy samples.I am still not completely happy with the quality of this version, so it will stay in the prototype phase. At this point, it is mostly a training data issue - there isn't enough good images to overcome the issues in the current version:<ul><li>AI sometimes draws Maya in long hair - in some way, this is a feature, not a bug.</li><li>Her military uniform keeps showing up. Unfortunately, in both the anime and the official arts she wears a single outfit, making it hard to break the linkage of the clothing and the character.<ul><li>I didn't use the time lapsed version of her in training - it is different enough and will only further confuse AI.</li></ul></li><li>Her expression feels stiff from time to time. I think the model might be overfitting a bit. Try lowering the weight to 0.8 if you see it happening.</li></ul>I might revisit this model in the future when I find a better method.Original notes for v1:Flux prototype of <a target="_blank" rel="ugc" href="https://civarchive.com/models/248428">https://civarchive.com/models/248428</a>, trained with the same dataset.Flux doesn't always get her hair length right - which isn't a bad thing if you want to see a long hair Maya. Otherwise you can specify short hair in the prompt.

Ibuki Maya - Neon Genesis Evangelion

Retrained with lower network dimensions to improve model stability and consistency. This is the first model that I trained with dim=2. It is very small, and turned out alright. It is reasonably functional, but still not as flexible as I want it to be. Most likely because there isn't enough good training data.

Files with SHA256: f985a31560ded48acd5ddf6ef2ae41af02ff4c99bbd0c7f6bf972f531768afcf