according to this my model: with over 100.000 steps should look like dog shit bc its overtrained. but the picture quality and extreme facial likeness says otherwise
well your lora is kinda a special case because most people would go the other way around (you trained at 10k images for i imagine 10 repeats, a lot of folks would do 10 images at 10k repeats), but my point still stands - that many steps will always get either artifacts at some point or diminishing returns
I only have 8gb of vram so unfortunately I have to stick with 512x512 resolution for now.
i also recommend watching robert jene's videos on lora training on his yt (probably best resource you gonna get, his explanations of settings with grid comparisions are extremely useful to see what actually does anything)
this! use high quality images! this does not mean high resolution but "clean" images, no filter, no distortion, no covered faces, no too far away subjects etc. also different fecial expressions of the same person makes a HUGE difference in correct facial features
I used his settings for my latest trainings and that helped my results improve a lot! I had to change some stuff due to hardware limitations but I'll take a look at captioning and my dataset, but I believe that the pictures that I chose are good for training, I try to focus on the subject face with clear images and different facial expressions.
from what i have found the overtraining artifacts usually happen from undertagged photos, so lora starts treating a bunch of untagged details as core features, hence the artifacts
Strange, I'm using the WD14 captioning method and only remove tags like 1boy and stuff like that. I don't know if that matters but I usually remove detailed descriptions of the background.
Yeah but I'll compare with your captions to see if there's something I can improve.
ai upscaling could work, but they might not be 1-1 (depending on the original quality it might change some features), keep in mind you dont need a ton of images as long as they are quality - 5 high quality images are better than 50 shit ones
Post your lora, maybe its undertrained
Thanks for replying.
It seems like you don't have xformers installed, to be honest I don't really know how to install it, I believe it was done automatically when I installed A1111, there should be some tutorials on youtube explaining how to do that.
i got the all nans unet exception when i overtrained my model (i did 15x40x6, last two epoch threw out the error, after i set the no half command line all i got from my lora was latent noise, im working on a 3060 12gb)
From what I've read, it could be a bug in the SD version 1.7 as well. When I train my model with WD14 captioning in Kohya I seem to get this issue. When I use Blip captioning instead to train, I can generate the images and don't get these errors.
When I use Blip captioning like you suggested in your tutorial, I don't get these issues so not sure why they happen with WD14 captioning. I got semi good results by just following your tutorial so would be great if you could share the training settings you're using now.
