This LoRA can either make a realistic scene into an anime style one, or make an anime image into a photorealistic one.
Works both in I2I and T2I.
Note: This is a concept LoRA, not a style LoRA. The model knows what an anime is, the LoRA just plays on these weights. The training dataset does not contain any anime images.
The sample weights range from 0 to 8 (full effect). The last sample is a anime image generated in T2I and made realistic in I2I with negative (-7) slider weight.
Hope you like it!
Description
FAQ
Comments (3)
What prompt did you use, because Klein can do this without any lora by default. The problem I encountered is that it's not as consistent and accurate as QIE, it usually changes hairstyle or keeps the image too stylized, for example, more anime'ish images won't become as realistic, with characters having heads too big. I was hoping someone would make a lora like this to fix that problem specifically, but I think a more general consistency type lora would be more useful, since you can already just prompt "transform into a realistic photograph" or "transform into an anime picture" and it works relatively well already
"an asian woman wearing a kimono in a zen garden", seed 9.
These LoRAs are heavily seed dependent. On some seed the effect applies immediately, on some other seed it'll take a heavier weight, plus the "style" also changes between seeds.
I recommend setting the weight to 3 then locking the seed and incrementing manually until you find the style you want, then play with the slider weight.
In my case, seed was lockes for all samples, so I had no issue.
You only have 3 of the same lora my guy. You need to train at least 3 more of the same lora.