Sign In

Dot on Forehead

6

98

330

2

Updated: Jul 7, 2024

concepttool loraforehead mark

Verified:

SafeTensor

Type

LoRA

Stats

98

330

86

Reviews

Published

Jul 7, 2024

Base Model

Pony

Training

Steps: 1,320
Epochs: 1

Usage Tips

Clip Skip: 1
Strength: 1

Trigger Words

blue dot on forehead
red dot on forehead
black dot on forehead
green dot on forehead
yellow dot on forehead
purple dot on forehead
orange dot on forehead
white dot on forehead

Training Images

Download

Hash

AutoV2
E78BEF182A
default creator card background decoration
Failed Generation Contest Participant
urusan's Avatar

urusan

This model is really straightforward, it generates a colored dot on a character's forehead. The trigger word is "COLOR dot on forehead". Explicitly trained on red, green, blue, purple, orange, yellow, black, and white but might support other colors. Note that you might not always get the target color for a variety of reasons.

That said, the point here isn't usually to make a colored dot but to form a "nucleation site" for some other brightly colored forehead feature to emerge around. The recommended usage is to keep it active for the first 15%-20% of steps and then turn it off. Automatic1111 has this built in with the syntax [black dot on forehead,::0.2] to keep it on for the first 20% of steps. If you have access to an extension to change the LoRA's weight, it's recommended to turn it off as well at the same time, so you can reduce the model's bias on your generation (though it will unavoidably cause some bias in the overall composition).

For best results, you should form both the dot and your intended goal object at the same time, so it's more likely to be vaguely shaped like the goal by the time.

This was inspired by issues I was having with my dragon character's blue forehead crystal being unable to "punch through" the overall bias not to have objects on a character's forehead, despite substantial attention paid to the area. This LoRA solves that issue handily.

I'm considering this first version experimental and I hope to see other people build better and/or similar LoRAs along this same vein. It can definitely be improved but it's good enough to release.