How to use Dreambooth and
free google colab to
to create a model
For the Deforum Diffusion
notebook
Hey everyone, so I have seen a lot about Dreambooth and one of my Patreon members encouraged me to tackle getting this
into the Deforum notebook and after watching some videos and experimenting a bit I managed to get a working model
into the Deforum notebook and it works great! It will require the following, you need to create a huggingface
account to get an access token and you will need some images to train the model on, whatever you want it can
be you, your friend, a dog, a human hand (Can we actually make realistic hands now with this technology?)
and it will take about an hour or so to create the model. You will do this on the Dreambooth notebook and the link is below.I did this on a free google colab account without any issues, I then copied the model over to my google pro account and used it in the deforum notebook.So this is a real cool feature, but also time consuming and the model wound up being almost 4 gb so this is not something you will want to do very often, I would kind of plan out what you want to do
and I used 512×512 images for the training. So this process is a bit of an undertaking just to warn you!
Ty for watching and please link your work in the comments! I love to see what people are creating with this!
—————————————————————
Here is the github link to Dreambooth
https://github.com/ShivamShrirao/diffusers/tree/main/examples/dreambooth
The Dreambooth Discord server
https://discord.com/channels/1023277529424986162/1023645424604692500
The Dreambooth notebook you use to create the model
https://colab.research.google.com/github/ShivamShrirao/diffusers/blob/main/examples/dreambooth/DreamBooth_Stable_Diffusion.ipynb
——————————————
Deforum notebook created by @deforum_art on twitter!
https://colab.research.google.com/github/deforum/stable-diffusion/blob/main/Deforum_Stable_Diffusion.ipynb
Deforum Discord server
https://discord.com/invite/upmXXsrwZc
——————————————————————–
Nerdy Rodent-If you run locally this is the channel to subscribe to!
He runs everything locally and has lots of videos about lots of great programs
and was one of the channels that introduced me to AI art programs.
https://www.youtube.com/c/NerdyRodent
———————————————————————
Science! (I love Science,some great channels)
https://www.youtube.com/c/EventHorizonShow
https://www.youtube.com/channel/UCEszlI8-W79IsU8LSAiRbDg
https://www.youtube.com/channel/UCZFipeZtQM5CKUjx6grh54g
https://www.youtube.com/channel/UCZFipeZtQM5CKUjx6grh54g
———————————————————————–
Synth V-The program I use to create the vocals for my song
https://www.youtube.com/c/DreamtonicsCoLtd/videos
Eclipsed Sounds-the maker of Solaria the primary voice bank I use with Synth V
https://www.youtube.com/channel/UC9_fAgkUrsjC4YiyKuwCPMQ
Emma Rowley-The singer who did the vocals for the Solaria voice bank
https://www.youtube.com/user/EmmaRowleyMusic
I will be making tutorials on this program very soon!
——————————————————————-
Serpents of the Everloving Light
https://www.youtube.com/channel/UC_e02YP9VatnN0gKEiu0YGg
——————————————————————-
Patreon Page where I post all my weekly prompts and settings for my images and videos!
https://www.patreon.com/prophetofthesingularity
—————————————————————–
Prophet of the Singularity Hyperfollow Social Media and Music Links
https://hyperfollow.com/prophetofthesingularity
Ty for Subscribing! More Tutorials, Videos, Studies and Music coming soon!!
Hey man! Just want to say thank you for sharing your progress in machine learning. I've gotten a lot of inspiration watching your videos.
is this google colab the exact same as renting a gpu and training a ckpt directly? because this converts the diffusers to a ckpt. So in your experience is it just as good?
Thanks
super fire thanks man
Great! I should be able to follow this. Thanks! Will try as soon as I find or shoot the right set of photos. I think I'm going to shoot for an AI renderd Christmas card this year. Father Ted was a great show. I went to visit the island of Inisheer where the intro was shot. The shipwreck is still there.
What would be the proper way to cite this Colab document on insta? Is there a page or yourselves I can tag? I'm able to make a shit ton of stuff off your tutorials as a launch point and wanna give credit where it's due!
you very appropriately picked a youtube brand name! thank you for taking the time to make these while you create!
how to train the model to have good results, regardless of the seed?
I've made more than 10 models, with different configurations, and all of them have inconsistencies in the results
I think that some of the code got updated again and the colab doesn't work, right now
very nice one！ ； ）
Thanks again man. You are really helping people out. On this one though I never got the notebook to work. Messed with it all evening. Seems like someone else in earlier comments experienced the same thing.
Thank you again for great videos ♥
you dont have to use the class in the promt? it was guy right? i still need to play with this.
i heard the training images need to be 512 x 512 can you confirm
you are doing God s work … great stuff
one more thing that confuses me that not many people cover.
so from what i understand you can only train one thing per ckpt file. So if you wan to do multiple trainings for say different faces you need to create a new ckpt file for each on and switch them. Does that sound correct. So if i have 5 faces i want to train to use. i need a check point file for each one and then switch between them.
This is my understanding. i will start messing with it tomorrow as i have a story i want to put together and i need consistent faces for each character.
Thank you for all the help btw! 🙂
Does this work for animations with a certain style? Or only objects/photos?
Dreambooth discord link is broken so it seems
i had not accept the model licence so i got an error if you run into this their is a link on the notebook that will send you right their.
i got an out of memory error "RuntimeError: CUDA out of memory. Tried to allocate 1024.00 MiB" I am going to try to run it on pro account
ok. so what i was doing was uploading the images myself but not in the correct path for the training so when i went to look for the training images it found nothing and errored out.
Did something change recently? Because I can't find the model.ckpt file in my Google Drive where it's supposed to be. Everything else is there except that one file.
So if I understand this correctly when you create your new checkpoint file, it not only contains all of the data for your chosen likeness, but also all of the previous training data contained within the sd-v1-4.ckpt model?
Hey, how to accept the model license ?! I cant find another tick box, I already clicked one box though, but it still asks me to accept the model license.
Do you have any recommendation on how to change the value in the training cell?
im encountering an error, a few it would seem. first off the section that is supposed to take the longest happened instantly. in the next section where it asked about weights_dir it gives me an error if i leave it blank (no idea what too add) if i do add something random the next section just gives me an error