Key moments:
11:33 Negative Prompt
12:51 Restore Faces
13:47 Image To Image
Enjoy this informal demo of Dreambooth, a stable diffusion AI model built by Google engineers and BU researchers. This software is completely free to use and open source with many distros still coming out! Please note, I trained my model with only 20 photos thus this is nowhere near the highest potential of the software. I’m also a non-technical person so if I’ve botched some of the terminology or explanations I apologize. Learning is an iterative process. Here are some of the links I’ve described in the video below:
Please consider donating on my Ko-Fi page so I can continue cranking out content like this!
https://ko-fi.com/virikemen
(The UI I used to work with the model)
https://github.com/AUTOMATIC1111/stable-diffusion-webui/
(The training environment I used to generate my model)
https://colab.research.google.com/github/TheLastBen/fast-stable-diffusion/blob/main/fast-DreamBooth.ipynb#scrollTo=1pH1oP-7yBZm
(A community-sharing platform for AI models)
https://huggingface.co/models
(The official site for Dreambooth)
https://dreambooth.github.io/
what kind of GPU do you have?
You actually don't want to oversaturate the model with too many images because it can be overcooked and destabilize all other tokens. The dreambooth we can use IS NOT the exact same thing as in the dreambooth paper. Please be aware of this.
We got you 👉😎👉
"Automatic eleven-eleven"
There are huge differences between the formers and full versions outputs. They're both improving daily
And get on the discord bruv!
If you do pictures of you with others it will fill the model in with you with others.
Every picture you add risks destabilizing the entire model. You can also update your face picture settings in the….settings.
same level of addiction for dreambooth, my pc couldn't meet the installation requirements so I run it with a google collab doc. I usually use around 20 pictures, I got some close results but manytimes I discover it creates people with different faces from there trained model
You thought crack was bad, try dream-booth
you should try working with negative prompt more
Can you train the model to understand how the back of you looks too? In every training example I have only seen frontal / and mostly waste up. Are any of your pictures full body?
1:35 count the number of knuckles agein 😉
16:25 you should turn on inpainting in full resolution, and turn the padding of that setting up to about 100, that will give you great results, when inpainting faces always use a the same width and height
bump up the steps to 50. I too have fallen down the SD rabbit hole
The big problem with AI art is that it is horribly ugly and repulsive, and it's not clear how that can be fixed.
You mentioned that you liked the dog image "getting a haircut" for the quality of the hands. I can only notice that the hand has five fingers plus a thumb. Weird. I've noticed this on other AI-generated images too. Something that still needs to be figured out, I guess.
I don't believe the AI got those hands – I suspect they did some photoshopping before posting those images. I don't trust Google for anything.
If you bump up the sampling steps, you'll get much better results. You were on 20, which is pretty low. However, I've noticed that too much can make it worse. I find that between 70-90 is a sweet spot – but it all depends on what you're doing.
The hand in "Getting a haircut' had five fingers and a thumb…by over all looked pretty decent.
Can I install this on Mac?
The pencil sketch gives you larger eyebrows. 🤣🤣