1902it

317 images
6 repeats

4 batchsize
3 epochs
1902it total

3.2e-4 learning rate
4e-5 text_encoder
3.2e-4 unet_lr
768 res
128 dim
128 network_alpha

wd14 swinv2 v2 interrogator
no vae 
0 keeptokens

got some interesting results with different learning rates on this one.
>multiply 8e-5 by batch_size for unet_lr. set unet_lr as learning_rate also.
>multiply 1.5e-5 by your batch_size for text_encoder_lr.
>batch size 4 has unet at 3.2e-4 and text encoder 6e-5.

training notes i took in the process of adjusting rates:

the original parameters i trained with were my default (lr 2e-6, text 5e-5, and unet 1.5e-4).
this as usual worked fairly well, though i wanted to mess around with learning rates to see if it would make a difference. for the next set, i used the instructions in the greentext above and adjusted to newer learning rates.
the main goal was improving the style while 

=v1=
novae		old		new		delta
lr = 		2e-6		3.2e-4		increase
text = 		5e-5		6e-5		increase
unet = 		1.5e-4		3.2e-4		increase	

observed changes:
black thighhighs turned white at 1.0
short hair in prompt ignored (turned long)
background more consistent with older (lr)


=v2=
novae		old		new		delta
lr = 		3.2e-4		3.2e-4		0
text = 		6e-5		5e-5		decrease
unet = 		3.2e-4		3.2e-4		0	

observed changes:
short hair is no longer ignored (fixed)
black thighhighs still turn white
background is closer to weight 0

=v3=
novae		old		new		delta
lr = 		3.2e-4		3.2e-4		0
text = 		5e-5		4e-5		decrease
unet = 		3.2e-4		3.2e-4		0

observed changes:
short hair is still fixed
white thighhighs are fixed; they are now black

this is the final version that is uploaded. in v1, the changes made to the learning rate incorporated more of the artists style while also not following the prompt as well; the thighhighs were white and the hair was no longer short. i figured this was due to the increase in all the learning rate settings overall from the original. to compensate for this, text_encoder lr was lowered from 6e-5 to 5e-5. this fixed the issue with the hair but did not fix the thighhighs. v3 lowered text_encoder lr to 4e-5, which seemed to do the trick. overall, the differences between the original version and the final one are slight, but it does incorporate more of the style while also adhering to the prompt, which was the goal.