tag:hoshiguma yuugi
1girl,single horn, pointy ears, blonde hair, long hair,yellow eyes,shackles ,white shirt,skirt,
Description
FAQ
Comments (5)
Hello. Now I'm studying to make great works like your LoRA.
Let me ask you some questions.
1. How many images do you have? Also, if you don't mind, please tell me the number of epochs and the number of repeats for reference.
2.Do you want the AI to learn fan art images as well?
3.What optimizer_type are you using?
4. I also love the Touhou series. who do you love the most?
Hello.
1. The touhou and genshin characters I made recently usually have 200 pictures, about 100 normal clothes, and 100 other clothes. Before making touhou characters, I usually have 40 to 60 pictures, all is the official costume. Regardless of the number of pictures, my epoch and repeats are always multiplied to 80, and the webui can see the lora training information.
2. In fact, I seldom use official images, most of my lora use fan data sets, and an official image is useless.
3. Most of the time I use the original version of lora training script in February, there is no optimizer_type option, I occasionally use the online version, probably AdamW8bit
4. I like Sakuya, especially Sakuya who is being bullied.
@kozue
Thanks for the answers.
1. so does that mean the total step count is roughly 16,000 ?
It takes quite a bit of time, doesn't it?
4. sakuya is cute. You have good taste. I like Remilia. Very beautiful.
@changpingjitian641 If there are 200 sheets, 80 times the number is 16,000 steps. If it is my computer, I will use 2 batches, which is 8,000 steps. If it is an online server, I will use 8 batches, which is 2,000 steps. The time is basically 20-40 minutes about.
I have observed some friends, they use the same parameters as me, but the effect is not good, mainly because of the selection of the data set, I will try to choose pictures with very different styles, even if it is a child's graffiti, as long as it Contains character traits, which I will use as well.
Thank you very much!
It was very helpful!








