A LoRA based on Latin streamer Arigameplays.

~ WARNING! NOT COMPLETED ~

This is my first LoRA.

It is not complete. It is in its first phase of training and I am working hard to significantly improve the results.

I post it to receive suggestions and improve the model together with the community. Please let me know your considerations.

At the moment we could consider that the model generates the distant cousin of the celebrity in question.

~ UPDATES ~

v0.2

• The eyes now seem to distort less frequently.

• More flexible

• It is still difficult to ensure that the images generated are faithful to the subject.

• It seems to work better for portraits than for cowboy shots or long shots.

recommended settings:

Low CFG scale. Example: 6
0.6, 0.5 weight
restore faces: on

~ LAST VERSION ~

v0.1

recommended settings:
0.6 - 0.8 weight

Feel free to experiment widely with the model.

Let me know your suggestions and considerations.

Working in progress...

~ TRAINING SETTINGS ~

(v0.1)

Trained on Colab following the Hollowstrawberry Guide

DATASET: 32 images (mostly from instagram)

10 repeats

20 epochs

flip_aug: off

shuffle_tags: on

batch_size: 2

unet_lr: 5e-4

te_lr: 1e_4

scheduler: cosine_with_restarts at 3

lr_warmup_ratio: 0.05

min_snr_gamma: on

network_dim: 16

network_alpha: 8

(v0.2)
DATASET: 54 images (mostly from instagram)

5 repeats

30 epochs

flip_aug: on

shuffle_tags: on

batch_size: 1

unet_lr: 1e-4

te_lr: 1e_5

scheduler: cosine_with_restarts at 3

lr_warmup_ratio: 0.05

min_snr_gamma: on

network_dim: 128

network_alpha: 128