Sign In

Doki (There She Is)

80
551
72
13
Verified:
SafeTensor
Type
LoRA
Stats
36
52
Reviews
Published
Jan 5, 2025
Base Model
Illustrious
Training
Steps: 2,020
Epochs: 4
Usage Tips
Clip Skip: 2
Strength: 1
Trigger Words
tsidoki
white eyes, white shirt, white skirt, short skirt
Hash
AutoV2
89A1C07FA2
default creator card background decoration
Bronze Character Badge
Smez's Avatar
Smez

Doki, a little white rabbit from There She Is.

UPDATE: This LoRA is so hard to make I make it to become my personal benchmark to see how I can push LoRA to the max with VERY...VERY limited training images in both quality and quantity.

It is this bad...

[Illustrious] NEW! Design closer to the source material.

Use tsidoki to summon her, use white fur, white eyes, white shirt, white skirt, short skirt to form her appearance, Omit some tags for a closer composition tag

vPony [NEW!] Holy cow, the generalization is amazing!

Use doki, white rabbit, pink bow, white shirt, white skirt, white shoes, white fur, body fur to summon her in her full glory. Omit some tags for a closer camera angle tag.

I retry the hardest LoRA I've ever made, with very few...and low-quality training images (consisting of ONLY a screencap of the flash animation) and this is the result, oh my god!

V2:

Train in the best result carefully selected and photoshopped from the 1st generation of LoRA
Still, this model still work best on NovelAI model. Others model will not give an accurate representation of a character.

Use tsidoki at the start of the prompt to summon her (no 1girl, no chibi), use white eyes, rabbit ears, body fur to form her head, use pink bowtie, white shirt, white skirt to summon her clothings

This attempt is looking better and gives me more consistency in each generation. Finally!

==============

V1:

This is probably the hardest LoRA I have ever attempted to make, takes me half a day to reach this point, and the result still does not satisfy me, but I'll upload it either way. Works best ONLY on the NovelAI model. others will not give an accurate representation of a character.

The source material is very limited in color, character blends in the background, too simple, and too unique, it's too bad I have to use well-generated pictures of the previous generation of the LoRA and feed it back to the newer training sessions.

Use chibi, tsidoki at the start of the prompt to summon her, use white eyes, rabbit ears, blush stickers, body fur, white fur to form her head, use pink bowtie, white shirt, white skirt to summon her clothings