Sign In

Kirei Kisegawa 黄瀬川輝麗 | Shinobi no Ittoki

81
676
0
Updated: Oct 5, 2024
characteranimefemale
Verified:
SafeTensor
Type
LoRA
Stats
312
Reviews
Published
May 30, 2023
Base Model
SD 1.5
Trigger Words
kirei_olis
Hash
AutoV2
ECDBFDC3AD
default creator card background decoration
Second Birthday Badge
JasonCMK's Avatar
JasonCMK

Kirei Kisegawa from Shinobi no Ittoki


Update v1b (30th May 2023):

  • Made minor adjustments to the dataset.

  • Fixed the training parameters.

  • Set the Noise Offset parameter to 0.05. Based on my limited experience, noise offset can help prevent a black outfit from becoming white, or vice versa. However, setting it to higher, e.g. 0.1, seemed to mess up the learning rate of D-Adaptation v3.

  • Retrained the model with D-Adaptation v3. (v3 is really interesting. I noticed that the learning rate occasionally increases on its own, even with Constant schedule, as observed in TensorBoard. In my latest attempt I decided to try Cosine schedule with NO warmups and NO restarts.)

  • Experimented with finetuning (https://github.com/kohya-ss/sd-scripts/pull/491). Not sure if it helps or not tho.


Prompts (v1b):-

  • Trigger Words:
    kirei_olis, long hair, blonde hair, side ponytail, green eyes

  • School Uniform:
    aqua scrunchie, black shirt, red sailor collar, neck ribbon, long sleeves, grey skirt

  • Outfit 2:
    helmet, bodysuit

  • Outfit 3:
    orange scrunchie, blue hoodie

  • Outfit 4:
    orange scrunchie, white shirt, sleeves rolled up, blue shorts

  • Outfit 5:
    aqua scrunchie, turtleneck sweater, brown coat, brown shorts, black pantyhose

  • Extra:
    1girl, solo
    (please refer to my sample images)

Weight: 1 (you may experiment yourself)


All my models were trained with 512x512 pixels.
For optimal results, please upscale your generated outputs with your preferred methods.


My sixth model.

For some reasons, I trained a model for a character that I wasn't really interested in, again. Btw, I tried using Lama Cleaner to clean up some of my dataset this time.

Hopefully some of you will like it.