Edit 10/23/23: Version 2 is now uploaded. Larger dataset means better and more consistent output!

This LoKR is designed to generate Indian/ South Asian looking faces and skin-tones when applied to SDXL 1.0 based Checkpoints.

LoKRs are specific kind of LoRAs that use some fancy math to keep their size very small (This is just 7MB, compared to 200-700MB sized LoRAs!)

What this does mean, though, is that:

1. The LoKR may not translate to other checkpoints as well as the one it was trained on (or at all)

2. The output may range from sublime to creepy, for the same prompt, based on how the AI influences your base checkpoint.

This was trained on SDXL 1.0, but seems to work best on the ZavyChromaXL checkpoint (https://civarchive.com/models/119229/zavychromaxl)

Put this in your models>LoRA folder, and use it as you would use a LoRA.

Settings vary based on checkpoint, but very broadly:

CFG: 7-8

Steps: 20-40

Strength: 0.3-1

Sampler: euler a, DPM++ 2M SDE Karras