Vibemix Color Style

Description

Experimental LoRA that combines a mix of different styles from a wide array of image sets with interesting color themes, dynamic poses, and lineart from both anime and furry settings. Trained with the focus of giving a more vibrant, colorful, and dynamic look to the prompt. Works better in a more painted or cartoon style. Might look too saturated on realistic prompts unless that's what you aim, for example, inpainting a certain colorful detail.

Use

Use it if there's a prompt that you wish had more color or a more dynamic and vibrant look. Depending on the model, it can be prone to oversaturation, so adjust the weight accordingly. I've been using strength 0.7-0.9.

Tag keyword suggestions:

Trigger word:
vibemix

Regarding Furry or Anime:

Exclude "furry" in the negative prompt if you want anime or humans only.
Include "human" if you want humans in your furry prompt.
Recognizes both Danbooru (e.g. 1girl, 2boys) and E621 (e.g. solo, duo, male, female) tag systems

Tag suggestions:

Any [color] tag brings out the LoRA. Reacts strongly with more vivid and saturated colors.
For example: "red shirt, orange shorts, turquoise hat".

Prompts that are usually associated with vivid colors, will also react strongly.
For example: "cyberpunk, vaporwave, glowing, paint, graffiti"

Forcing the LoRA to trigger on more details, for character customization:
For example: "tattoos, [body_part] markings, glowing markings, glowing [body_part], stripes"

Parameters

Depends on your model and preference, but I recommend:

Sample: Euler a
CFG: Lowering 1 or 2 levels might help if it's too saturated.
Strength: 0.7-0.9

Known issues:

Haven't done much testing yet, but I've found these issues when running the LoRA as opposed to running the model alone:

Oversaturation might occur with:

• High LoRA strength -> reduce the weight

• Too many vivid colors in the prompt -> replace one or some of them with less vivid colors and the overall saturation should lower (e.g. turquoise/yellow 'glows' more than blue/orange)

• Models that are already trained on vivid images -> Change model or reduce the weight, CFG, and use less vivid colors (e.g. blue instead of turquoise)

Color bleeding might occur:

• When two vivid colors are adjacent they could mix and produce a different color (e.g. green hair + turquoise hat might result in the hair and the hat to be of the same color that is a mix of the two) -> Prompt with colors that contrast more, or simply continue to generate with different seeds. Sometimes it happens randomly.

• A color might bleed into another object that has a different color named in the prompt -> changing the object name for something broader sometimes works (e.g. blue pants -> blue legwear)