Key takeaways from the Dreambooth paper for better Stable Diffusion results ft. Hasan Piker
HTML-код
- Опубликовано: 28 май 2024
- I wasn't getting good Stable Diffusion Dreambooth outputs so I read the original Dreambooth paper and this video summarizes the key takeaways.
Key Advice:
1. Use an uncommon 3 letter sequence for your identifier
2. Use 40+ images of the subject and 4000 training steps
3. Use 128 "class" or "regularization" images
------- Links -------
Discord: / discord
Paper: dreambooth.github.io/
Colab Notebook: colab.research.google.com/git...
------- Music -------
Music from freetousemusic.com
‘Sunset’ by ‘LuKremBo’: • (no copyright music) j...
Music by Bensound (2nd and 3rd songs):
Summer: www.bensound.com/royalty-free...
The Lounge: www.bensound.com/royalty-free...
#stablediffusion #aiart #techtutorials #tutorial #tutorials #hasanabi - Наука
Bahaha Hasan, a land owner in America haha. This is so great haha
Thanks for the precise and simple summary and hints
Glad to be one of the first commentors on this dope masterpiece of a video!
damn, that's nice as hell
Thank you guy! It's really give me a help
Love your straight to the point videos!
dude/sister, nothing fills me with more rage than clicking on a tutorial video and having them talk about how their day went
what's the better prompt parameters set up?
1. Keep both a context (a photo of) and the names (instance name / classname)
--instance_prompt="a photo of sks dog" \
--class_prompt="a photo of dog"
2. versus keeping only the names
--instance_prompt="sks" \
--class_prompt="a dog"
3. versus keeping a context (a photo of) but excluding classname from the instance name
--instance_prompt="a photo of sks" \
--class_prompt="a photo of dog"
What's your observations?
Number 1 is the recommended way to do it, but i don't think the recommendations are rigorous
Nice didnt realise the three character limit helped...nice, Heres a question I'm running 50 images of an object 500 reg images and 5000 steps of training. Even at 50 training images would you consider not upping the reg images...and what of the steps is 5000 too much you reckon?
6000 made everything over-cooked, so 5000 is... risky maybe??? I found that upping the reg images made training take so long it was too annoying to deal with. I would just stick to 4000 and 128 lol.
@@lewingtonn just off to bed so I've stopped training just as I got your message...thanks ill make those changes and fingers crossed it will create more coherent outputs 😋
Thanks man. Very helpfull and entertaining videos even though we are all serious people here :)
can i train logotypes or some form of graphic design with this?
Honestly I have no idea... you can definitely do models that reproduce physical objects like bags and mugs
Works fine on my end, sir. Perhaps your collab sucks, and you should step up, run it locally and try to understand the technology fundamentals...
which works? The 5 images with a smaller amount of training steps? I'd be really interested to hear what you did!
@@lewingtonn i ran it locally with the SD 111111 deploy used 12 imgs in different lighting settings
As in, you TRAINED it locally? Can I ask how many training steps and regularisation images? I want to try and replicate your results if I can because 5 images definitely didn't work for me earlier
@@lewingtonn correct, only a thousand steps cause my stoned ass forgot to change the setting. Dm me for my twitter , I'm happy to help you out
Regularitation, a shitload ...don't recall the exact number. Should look it up