Character from Dracu-Riot! by Yuzusoft.

Training settings:

20 Epochs (this version is 14th epoch checkpoint), 20 steps, fp16, Lr=1e-4, cosine scheduler, 10% warm up, AdamW, CLIP skip 2

For using this LoRA, please refer the sample grid for CFG and LoRA weight.

For simplicity, use CFG scale within 4-6 and LoRA weight around 0.9.

This LoRA is trained on https://civarchive.com/models/17394/quinns-colorfulmix

Sample prompt used for all samples (may have small adjustment):

Positive:

<lora:Elena_v4_2:0.9>, Elena_Owen, masterpiece, (masterpiece, best quality), a girl, solo, (mature female:1.2), HDR, Ray Tracing, glistering shiny, best colors, bloom, (depth of field), (looking at viewer, smirk), (low_angle), beautiful face, beautiful eyes, ((silver hair, thick hair, black ribbon)), (black panties, pantyhose, cameltoe), (topless), (glasses), parted lips, boobs, (beautiful eyes, light blue iris), squatting, pussy, (inside, windows open, office, sun ray, sunlight, framed paint)

Negative (using https://civarchive.com/models/7808/easynegative):

easynegative

CFG:4, DPM++ 2M Karras, 20 steps, R-ESRGAN 4x+ Anime6B upscaler, 0.45 denoising strength, 512*512 generation size and upcaled by 2. Finally using R-ESRGAN 4x+ Anime6B to upscale again for 2k image.

P.S. I realized early samples used Elena_Oven instead of Elena_Owen, somehow still worked fine. I guess those two words share very similar latent embeddings.