Anonymous
7/8/2025, 1:30:38 AM
No.8654067
>>8654064
Yep, there's 2-3 images without the aura and wings in the dataset. It's all the same artist though, so style bleed is pretty bad.
>>8654065
Most pics I don't have in the dataset are part of a variant set, so no point to include 10 same images with minor differences. These are all the same artist (official creator), I've found like 2-3 other fanarts but they are complete crayon tier and changes the design so are a no-go for the dataset.
I'll likely retag everything again manually, train an almost fried lora and then start introducing artificial examples into the dataset with different artist styles. I have the metadata in my pics, it's in the stealth png format (works with the extension), but if you don't have it I used "hetza \(hellshock\)".
Yep, there's 2-3 images without the aura and wings in the dataset. It's all the same artist though, so style bleed is pretty bad.
>>8654065
Most pics I don't have in the dataset are part of a variant set, so no point to include 10 same images with minor differences. These are all the same artist (official creator), I've found like 2-3 other fanarts but they are complete crayon tier and changes the design so are a no-go for the dataset.
I'll likely retag everything again manually, train an almost fried lora and then start introducing artificial examples into the dataset with different artist styles. I have the metadata in my pics, it's in the stealth png format (works with the extension), but if you don't have it I used "hetza \(hellshock\)".