[-Follow the settings-]❇️
Absolutely! I am currently working on an article for Civit and Huggingface, where I explore the concept of low dim and low-number of images. I had an idea that it would be really interesting to challenge people to experiment with this setting and showcase their results in the article as well.
(a character that i trained with this setting)
-RULES-
some settings is [strict], that means that you can'not change!
some settings is [suggestive], that means that you can change if you want.
a proof of your dataset can be good for some users, but optional.
only concepts and characters/person, style is kinda bad at this config, if you manage to make this work, that will be a extra point for you.
The theme can be ANYTHING that you want, just follow the setting and the rules.
if you're making a preview, i suggest you to use 3m sde exponential.
if you're making a dataset, will be a good idea to remove the background from the subject, because it can be easily learned.
-THE SETTINGS-
# If your trainer don't have prodigy, !pip install -U prodigyopt
[strict]
Dataset number of images: 4/5, nothing more, nothing less, don't worry if you want to train a celeb face or a person, that's works.
optimizer_type = "Prodigy"
learning_rate = 1
optimizer_args = ["decouple=True", "weight_decay=0.01", "d_coef=2", "use_bias_correction=True", "safeguard_warmup=False", "betas=0.9,0.999",]
network_train_unet_only = true
[suggestive]
dataset_repeats = 8
max_train_epochs = 20
lr_scheduler = "cosine"
lr_warmup_steps = 20
disable noise offset/multires.
mixed_precision = "fp16"
Others options is by our choice.
But, i can train faces with that settings???
yes:
the dataset of this lora consist on 4 images, no background.
also, one of the images is slight darken for better composition.
so, good luck!!!
also, if you still don't have any ideas for lora, make one of your favorite movie actor, everyone has a favorite movie, right? so take a screenshot of some scenes of this character, crop, and remove the bg, but follow the 4/5 images rule.