This LoRA does not accurately generate her main outfit.
Turning down the weight strength may be necessary when using several trigger tags.
Description
Comments (4)
I wish interested in doing a aranea highwing lora but I'm anxious with captioning her complex body armor and spear. With this lora, how did you go making it? Did you just use screenshots from the kingslave movie or mixed it with concept art, fanart and synthetic images.
It's difficult to give advice because only the dataset and the training parameters seem to really matter. Overfitting is a problem with the turbo model especially because it makes LoRAs aggressive. That can happen because of too many similar images in the dataset.
How much captioning matters, for LoRAs specifically, is difficult to say. Having a character trigger is probably good, using a VLM (e.g. Gemma 4 or whatever you want) is probably better than anime-style booru prompts, but it won't fix dataset issues from what I've seen.
I mix in some unrelated simple images at low repeats like 3D landscapes and generic female images to try to avoid overfitting, but I'm not certain how helpful that is.
With Aranea I'd be worried about the complex spear next to her complex outfit. Maybe use a good variety of images with and without the spear, and maybe mix other images that won't affect her face likeness or pollute too many tags from your main dataset. And hope for the best.
My LoRA doesn't recreate any outfit well because there are few good shots in the movie and she wears different clothes throughout. And I didn't want to deal with the problems above by adding concept art.
@dingus I struggle to even even get gpt image 2, grok imagine and seedream 5.0 to make synthetic images of her in her default outfit accurately. She definitely is going to need to be trained on rank 64. Thanks for the reply.
Would love to see Iris and Luna too. FF15 had great character designs, it's a shame the game wasn't received better.