V2.0I did not like that v1 of the lora skewed towards a certain set of training images from cosplay of the character. The new version has a much larger data set to train from and I experimented with training parameters until I got something that I liked. New images are using a variety of models such as pastelmix, anypastel, AOM2 nsfw, kenshi. Some images use another lora to change the style and you can see what it was from the image data, all of which you can find on civitai. USE HIGHRES FIX, it works without it but it has trouble getting eyes that aren't blurred kiss-shot acerola-orion heart-under-blade LoRa Tested only on AbyssOrangeMix2_nsfw and lora strength @0.7 Trigger words not really needed but it helps to use tags for her red dress, for example (blonde blonde hair, pointy pointy ears, red dress, black ribbon) (kiss-shot acerola-orion heart-under-blade) All outputs shown are using highres fix 20 steps / 0.5 denoise strength / Latent (nearest-exact) / Hires upscale: 1.8 | 
发表于 2023-6-7 19:57:25
