Sample images generated with Anything-V5 and also, EasyNegative was used
Suggest weight : 0.6-0.7
Trigger word : nanjo mao,red hair,short hair,hair between eyes,one side up,brown eyes
Outfit: hair ribbon,bow tie,school uniform,pink jacket,miniskirt,blue checkered skirt,socks
Description
FAQ
Comments (9)
I was thinking of making the same character, now thanks to you I don't need to do it XD, that saves me a lot of time.
did you only use the pictures from the anime or did you also use Sune's illustrations?
Also, how many epochs and steps did you use?
Hi there
>did you only use the pictures from the anime or did you also use Sune's illustrations?
both of them
>Also, how many epochs and steps did you use?
epochs : 10
steps : 75
Hope it helps
@doyagao
75 its to low, can you do this calculation and send the result?
(number of images x number of repeats x epochs)/ batch_size = ¿?
usually a result between 1500 and 2000 or a little more gives good results for characters when you have more than 15 images. (only if you are using Kohya trainer)
Also after testing the model I realized that it has some problems that are easy to solve with better Tags in the .txt files like:
- changing the tag "bow tie" to "neck ribbon".
- changing the tag "brown eyes" to "orange eyes".(depends on the predominant color of the eyes according to the image)
- add the tag "collared jacket" and "collared shirt", because when you add only the tag "school uniform" the ia tries to create a "serafuku".
A definitive solution to the serafuku problem is to add an image of the character from behind to the dataset
could you try it and see if it gives a good result?, no problem if you can't do it, you've already done a very good job.
PD: sorry if I sound too rude, when I translate from Spanish to English it sounds like that sometimes XD (mucho texto)
@LacraMaxima
>(number of images x number of repeats x epochs)/ batch_size
number of images (37) x number of repeats (75) x epochs(10) / 2 = 13875
and of course I'm using Kohya trainer.
also , I appreciate it for the advices of tags.
You are not rude and I enjoy talk with you. I'm not a English speaker too :)
@doyagao now I got very confused after seeing again the 75 XD, the formula I gave you is to calculate the total number of steps, you probably need to read this to understand what I mean.
@LacraMaxima
Sorry for the misunderstanding XD
Now the answer will be:
150 steps for each image
10 epochs
batch size was 2
@doyagao after sending the message I remembered how to read the metadata of the lora XD, then I saw there the number of total steps
the final formula according to the meta data is: (15x150x1)/2=1125
very different from what you showed me before XD, but maybe it's because of the method you used, mainly I use the Dreamboth Lora method.
@LacraMaxima
Oh yeah. metadata is there!
It's strange because I used 37 images to train and re-trained for several times
why it shows 15x150x1
@doyagao
yes, it seemed strange to me too, so I made a comparison between two Loras of yours and two of other people to see the differences.
here are the results:
https://colab.research.google.com/drive/1W3OqRjmh75mpk6GJINQrp6TpLY_8hBEd#scrollTo=eRpHO0t-3xoY
it is strange, could it be that you do another process after having trained the Lora, another thing could be if you train with the Fine Tune method, because I have no idea if it saves in the same way the metadata in comparison with the Dreamboth Style method, or if it has some option to not save the metadata.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.