Sign In

SECONDARY-CORE (Orange,Green,Purple)

28
141
34
7
Verified:
SafeTensor
Type
LoRA
Stats
141
34
Reviews
Published
Jun 22, 2024
Base Model
SD 1.5
Training
Steps: 11
Epochs: 858
Usage Tips
Clip Skip: 2
Trigger Words
Orange
Green
Purple
Black
White
Hash
AutoV2
A4B078D98D
default creator card background decoration
Failed Generation Contest Participant
AI_AIDIA's Avatar
AI_AIDIA

After testing out a quick lora with kohya-ss, I've tried to continue my series over there. The result is very unstable but it still serve it's basic function.

What does it do?
[COLOR+STYLE] It generates images with the SECONDARY color hue. It also ends up with a less obvious style if left unprompted. I've attempted to control the skin hue(or shadow) but the result is unstable and highly model dependent.

What are the triggers?
The main color triggers are: Orange, Green, Purple, Black, White.
The colors can be customized by playing around with the combination, the position of the prompt affects the priority of the color, the color in front will be prioritized. Check my prompts for examples. If you use other colors, their hue may shift closer to SECONDARY.
Optional add-on prompts: ,1girl, solo, portrait, (color) background, (color) skin
*There is also a SECONDARY prompt BUT I still don't know what it does...😢 If anyone know, lmk.

Recommended Weight
Depending on the models used, the weight is good from 0.3 to 1.2.
On average, 0.7 to 1.1 have a more stable result.

Recommended model:
RE-LIVE_VIVIDPAST makes the colors POP if you want a burst of colors.
All models will have different brightness results. Base on tests:
- Most realistic model have duller colors.
- Most semi-realistic model have darker colors.
- Most chibi model have brighter colors

CORE Color Series:


Reflection: Wanted to use way lesser data since more data didn't equal to better result, but I was not getting what I wanted so i slightly increase the dataset from 40 to 80. The color choice of the dataset fluctuated too much, it was not good for the lora. The final colors isn't stable when trialed on different lora. I will need to revisit this lora and improve the data set.