USE AT THE BEGINNING OF PROMPT:
score_9, score_8_up, score_7_up (anime coloring, 2d, anime_screencap), source_anime, 2d, original anime style,<lora:violet_evergarden-09:0.8>, violet_evergarden, 1girl, solo, long hair, blue eyes, blonde hair, ribbon, hair between eyes, hair ribbon, braid, red ribbon, violet evergarden,Weight Recommendation 0.8-1
Description
FAQ
Comments (8)
Sadly I couldn't get her mechanical Arm, I will try to do it again and see If I can capture it better.
The LoRA seems too burnt...
Typical condition of a burnt lora: too bright generations...
Really? I thought it's usually when you see the generation you get like a green burning tint as it's generating and also comes off rough in the middle of a generation if you look at the edges of the character and surrounding areas.
It could be because I ended up using Adobe Lightroom to remove the white haze fim that some anime have.
I'll check then and provide a lower epoch one then just in case :)
I checked my dataset and my dataset is generally brighter, which means I brightened the dark scenes of the anime. I had 350 images with 750 training steps and used prodigy with the recommended settings.
@Rule34Diffusion in general pony needs much less images to learn the concept. 30 images with 1 repeat 10 epoch is plenty enough = 300 steps per outfit/chara
If your images ended up too bright, then its bad. You can try this guide to get rid of the white haze: https://civitai.com/articles/5275/using-chainner-to-batch-image-process-dataset
Plus i also like to add one more step to that which is lowering gamma to 0.9 so that images don't end up too bright.
@LittleJelly ohhh I like chainner! I keep meaning to get it back since I didn't use it yet when I had my older computer.
but yeah I have noticed this with the image count, this was was much more than usual. but most of the time I go with 30 - 100 images haha xD
I even tried 2 images for fun and it actually kinda worked, but obviously not good because the character would go back to the pose in the images haha. I did try using that model and then train another one with a normal count and merge them(the 2 images was promo images, and the other was in-game extracted framers)
@LittleJelly I'll definitely do that then, and good to know thank you :D haha