Verified: a year ago
SafeTensor
Type | |
Stats | 793 |
Reviews | (141) |
Published | Jun 29, 2023 |
Base Model | |
Training | Steps: 5,750 Epochs: 10 |
Usage Tips | Clip Skip: 2 |
Trigger Words | IsokazeKC |
Hash | AutoV2 A1E5FF3FBB |
An experimental character LoRA of Isokaze from Kantai Collection.
Info
Training with 164 images
Followed the manual form [ [Guide] Make your own Loras, easy and free ],
You can access to this guide in Suggested Resources section.
Recommended set up
Sampler: Euler a, DPM family series
Sampling steps: 20 or 25+
CFG Scale: 7
CLIP Skip: 2
*Do not use [restore face] on anime character!*
LoRA Weighting: [0.9 - 0.8], (1 - 0.6)
Baking Step:
0. Drafting (512 x 512 res to find your desired sketch)
1. Hires fix (txt2img: 768 x 768, 768 x 512, 512 x 768, or above res)
2. inpaint (img2img: refine some unwanted effect, [Euler a] with 50+ step)
3. Upscaling (img2img: res*2 with [ControlNet Tile] & Ultimate SD upscale)
Upscaler for [Step 1 - 3]: [R-ESRGAN 4x+ Anime6B, 4x-AnimeSharp, 4x-UltraSharp]
I always skip Step 0, 2 :P
Basic prompts
IsokazeKC, serafuku, long sleeves, (grey) pleated skirt, (blue) sailor collar, (yellow) neckerchief, (asymmetrical legwear, single sock, single kneehigh), (white) gloves, tress ribbon
Some additional tags that you may use for reference
-ponytail, scrunchie, kappougi, apron, paper fan, uchiwa, cooking, grill, smoke, fish
==>NEG: (yellow) neckerchief
-two-vertical line serafuku, red eyes