Clothes are all tagged for easy swapping, compare the standard outfit picture versus the jeans & t-shirt for prompt difference.
~Krea2~
Portrait shots should not need face detailer for eye symbol but full body or far shots will need the extra pixels to form the shape.
Mask Dilation: 10
Inpaint Mask Blur: 12
Only masked padding: 80
Denoising Strength: 0.25
~Illustrious~
Run without eyes tags if you don't have a face detailer.
I tried training every viable booru eye tag.. To make the eyes work : run the face detailer with "symbol-shaped pupils, flower in eye". This version has eye-focused masked training but the pixel assignment isn't high enough via natural generation without the face detailer to replicate the symbols well.
Description
Krea2 V1
FAQ
Comments (2)
great Lora. I've observed that the Krea2 version has higher detail compared to the IL version. Is this a reflection of the difference in model capabilities? I tried training Ronova's Anima Lora, but the eye details are very coarse. Could you explain what eye-focused masked training is?
The eyes were masked, upscaled, and trained by themselves at a slightly lower repeat rate than the actual pictures. I found so far that Krea2 can train more details but I feel that it sort of lacks anime style tags/traits especially on the character. I do think the ceiling is higher and we'll see it rise for anime/3d/animation as Checkpoints like Jank2 Turbo hit.