Might be server strain due to getting popular; I remember submitting mine and getting something back in two minutes.
What's hilarious is how SD ate shit over the Taylor Swift deepfakes thing, when most of the images were generated using Microsoft's tools by abusing their shitty filtering. There were dozens of threads on /g/ and other parts of 4chan about how to evade their NSFW limitations. Most of the holes seem to be patched now, but people spent months generating images before that, and those images are still getting reposted.
Besides that, there are more niche models like Pixart and Lumina. It seemed like people were holding off on starting anything new, waiting to see how SD3 would turn out before investing a bunch of time and money in an alternative. Now that Stability seems to be collapsing, there is space for new entrants to try to become the next big thing in open source.
It’s still a new industry and they’re trying to figure out how to monetise. You can’t charge a subscription if nobody knows who you are so you offer a free service and mine data through Google to make a bit of cash.
Instructions on how to download and install Dreambooth can be found here. I have trained many LoRAs using the kohya-ss utility, it I’ve never touched Dreambooth so I cannot be of much assistance there. Fortunately, there do exists tutorials all over the place online.
Are you training SDXL? What res? That's way too high a time, especially for a 4090. I can train an SDXL LoRA at 1024x1024 with 50 images on one repeat for 50 epochs in about 2 and a half hours on a 4070. Are you accidentally using the dreambooth tab instead of the LoRA tab?
I'll try training stable diffusion 1.5 with lora since everytime I add dreamsbooth into the extensions folder, it stops the webui from working.
I have reference images, now how do you two feed the model said images with Lora?
Did you turn xformers on, and what are your network dim and alpha? I usually train in the range of 3,000-4,000 steps and it takes me roughly 30 minutes to bake a LoRA.
Your first step will be tagging the images for the Lora. Automatic111 should have a tagging module that will generate tag text files for each image. If all of your images share a similar tags you may wish to remove the similar tag, and use a singular trigger tag.
My workflow is using LoRA_Easy_Training_Scripts for the training and TagGUI for tagging and organizing images. I don't think anyone really uses Dreambooth anymore. If you're confused about config settings, download a LoRA you like from CivitAI, look at the metadata, and copy what they are doing.
Are there youtube videos, resources, etc...you have used to learn what you've given me for people to follow using Kohya? And theres alot of information to go through if say someone wants to train their models using Lora from the beginning just to get started.
I hope you've been pulling from AUTOMATIC1111's GitHub repository semi-frequently to keep it updated.
If you listen carefully you can almost hear a boomer voice in a board room saying "Everything's AI now, we need to get on top of that STAT. Do young people still like that rainbow stuff?"
Looks like they did another with a fat japaneese woman:
Do you use comfyUI or automatic 1111?
