This is an embedding of female character that I've generated or merged by embedding merging extensions thru SD based on 3rd-party asian doll likeness loras.
Since this refered person doesn't exist in the real world, any resemblance to anyone in the real world is unintentional and coincidental.
Description
Train with base model v1-5.ckpt，200 steps , 600 epochs.
FAQ
Comments (7)
Does training with 600 epochs take a long long time?
It only took 20 mins on my RTX 3060 12G display card. With Batch:3, Gradient:20(number of photo). 600 epochs looks as a big number, actually the training speed depends on how many steps you set. Batch:3 means in every step, there're 3 epochs went through. Training settings thanks to https://civitai.com/user/JernauGurgeh
@AceNeo
I'm a newbie so excuse my ignorance, but with 600 epochs in Kohya aren't you left with 600 safetensor files so what happens next with all of them? Can you share your configuration file for your Lora training? Training 5 epochs on my 4070 and config takes the best part of an hour.
My JSON config file is from here-
@FrozenGunner I got it now, basicly we're just talking about two different things here. The model type I created is called Textual Inversion(TI), AKA embedding file, followed with .pt as filename. And the model type you're talking about is LoRA, often followed with .safetensors as filename. The two kinds of model is trained very differently. LoRA file is trained using Kohya's GUI, and TI file is trained using Stable Diffusion Web UI. Since I only do TI training, I've never done LoRA training. I am completely unknow to LoRA training, so nothing I can do to help you man, sorry. All I have is TI training settings.
@AceNeo
Massive thanks for that clarification. I tried to train a Lora in Kohya today with no success... a broken arm in a plaster cast. Are there any tutes or tips you can give me about TI training using the Stable Diffusion Web UI, local install? My main interest is that TIs tend to not affect Loras used and change the character.
@FrozenGunner I like TI file because: 1st its file size so small, 2nd everyone who has Stable Diffusion is able to train it, 3rd it's easy to deploy. Since Stable Diffusion Web UI is easy to get as below.
https://github.com/AUTOMATIC1111/stable-diffusion-webui.git
recommanded work process is below from JernauGurgeh
@AceNeo
Thanks a bunch!
What do think of this...
https://www.youtube.com/watch?v=MLz0iM0M7Fk
Olivio says 2.5 hours.... fuck that.