Noah | NIKKE - v1.0 Review

Hey, I have tried your LoRA, not bad!

I do have a few suggestions based on the metadata,

First would be removing _ from tags in the dataset, replace with just a single space so noah_\(nikke\) becomes noah \(nikke\)

If you are using the standalone dataset tag editor you can change the default behavior in the settings to use space instead of underscore.

For existing dataset you can batch edit search replace entire caption _ to space.

Removing tags like 1girl and solo is generally advised for character LoRA.

Removing tags related to innate features of the character is also a good idea, i.e. pink eyes, pink hair so you do not need to prompt them.

Dropping your batch size from 5 to 3 will make it slower but is also generally recommended for better quality.

Edit: You should also train LoRA using the NAI leaked model instead of a descendant/mix, this ensures best flexibility with most anime models.

If you want to talk LoRA, SD, and/or NIKKE you may consider join me on discord (richyrich515) https://discord.gg/uk5qHgE62B