Trained on 308 dark skinned / tan only m-da_s-tarou images.
Images trained and gen'd with nai
Seems to need 'black rock shooter, tan, m-da s-tarou' to gen tan images
Description
FAQ
Comments (5)
Hi, what settings did you use for training? I would really appreciate a guide for style training in Lora. I have a dataset of 4k images and I would like to train the author's style, but I can't find the parameters for training
I normally use 5_name for my folder when training image sets that have sufficient data.
upon opening run_popup.bat
I always do the following
load json config file (no)
base model I use anyv3 but most recommend nai havent compared so idk
select your image folder
select your output folder
I dont save json (no)
regularisation (no)
continue (no)
batch size (cancel)
epoch's 4
dim (cancel)
alpha (cancel)
resolution (cancel)
rate (cancel)
encoder (cancel)
unet (cancel)
scheduler (cancel)
save epochs (yes / ok) set to 1 or press (cancel)
shuffle (no)
tokens (no)
warmup (no)
change name optional / (no)
comment (cancel)
4k images is a ton of images imo but who knows.
I recently discovered on some styles im training that CFG has an effect so im messing with that but it only seems to effect really blocky/random styles vs some I had already done it has no effect.
The less images you have the more repeats you want so 5_name -> 10_name etc until it trains
I don't prune my prompts .txt files I just use what comes with Grabber / Tagger.
Overall it's really easy & simple to do.
I normally always take epoch 3 / 4 from my output as they always come out the best imo.
@Numeratic Thank you very much, I will try)))
hi, sorry for a dummy question. Did you follow any tutorial that you could possibly link? I am turbo new and wanted to make some LORAs myself
https://github.com/derrian-distro/LoRA_Easy_Training_Scripts/releases/tag/installers-v1
I normally use 5_name for my folder when training image sets that have sufficient data.
upon opening run_popup.bat
I always do the following
load json config file (no)
base model I use anyv3 but most recommend nai havent compared so idk
select your image folder
select your output folder
I dont save json (no)
regularisation (no)
continue (no)
batch size (cancel)
epoch's 4
dim (cancel)
alpha (cancel)
resolution (cancel)
rate (cancel)
encoder (cancel)
unet (cancel)
scheduler (cancel)
save epochs (yes / ok) set to 1 or press (cancel)
shuffle (no)
tokens (no)
warmup (no)
change name optional / (no)
comment (cancel)
CFG has an effect on blocky/random styles vs some I had already done it has no effect.
The less images you have the more repeats you want so 5_name -> 10_name etc until it trains
I don't prune my prompts .txt files I just use what comes with Grabber / Tagger.
Overall it's really easy & simple to do.
I normally always take epoch 3 / 4 from my output as they always come out the best imo.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.