Sidebar Sidebar
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.

Run Stable Diffusion on your M1 Mac

S

senttoschool

macrumors 68030
Original poster
Nov 2, 2017
2,674
5,661
189539684-222482fb-63f7-4799-bfc1-005b84508e35.png
Stable Diffusion uses AI to generate images based on text: https://github.com/CompVis/stable-diffusion

Run Stable Diffusion on your M1 Mac’s GPU

How to run Stable Diffusion locally so you can hack on it
replicate.com replicate.com

Edit: One click installer available but with less features: https://github.com/divamgupta/diffusionbee-stable-diffusion-ui
 
Last edited:
  • Like
Reactions: eightace, Pressure and TarkinDale
One really cool thing about Apple Silicon is the unified memory. Normally, you need a GPU with 10GB+ VRAM to run Stable Diffusion.

But because of the unified memory, any AS Mac with 16GB of RAM will run it well. For example, an M1 Air with 16GB of RAM will run it.

However, to run Stable Difussion on a PC laptop well, you need buy a $4000 laptop with a 3080 Ti to get more than 10GB of VRAM.

I hope this trend continues and AS Macs become high valued machine learning inference/training computers.
 
This is great!
Using Dall-E already but will now be able to compare Stable Diffusion.
 
senttoschool said:
One really cool thing about Apple Silicon is the unified memory. Normally, you need a GPU with 10GB+ VRAM to run Stable Diffusion.

But because of the unified memory, any AS Mac with 16GB of RAM will run it well. For example, an M1 Air with 16GB of RAM will run it.

However, to run Stable Difussion on a PC laptop well, you need buy a $4000 laptop with a 3080 Ti to get more than 10GB of VRAM.

I hope this trend continues and AS Macs become high valued machine learning inference/training computers.
Click to expand...
I have a desktop 2080ti with 11gb VRAM and I can't do anything larger than 512x512
unified memory makes so much sense, having a separate memory pool is so stupid. I wish the desktop PC market could move to SOCs rather than separate GPU and CPU.
 
  • Like
Reactions: Xiao_Xi
Read about this and found it incredibly cool. Possibly stupid question however. According to the documentation 16GB is recommended but 8 gb "works". Does that mean its gonna swap my drive to hell or does it actually know when mac os says "hold on there buddy! I only got 8 gigs!"
 
SD on M1 max 32 core will generate a standard output image in about 30s. On an nvidia 3090 GPU we are looking at 5 s or so. Would be really interesting so hear if the 64 core M1 Ultra is in the 15s or of this is another task where the ultra chip doesn't scale well.
 
I just tried this out (a pain to get it all working, but that's beside the point). It seems that the rumours are true: Faces are the stuff of nightmares. Don't try to get it to make people if you want to sleep tonight! :)
 
innerproduct said:
SD on M1 max 32 core will generate a standard output image in about 30s. On an nvidia 3090 GPU we are looking at 5 s or so. Would be really interesting so hear if the 64 core M1 Ultra is in the 15s or of this is another task where the ultra chip doesn't scale well.
Click to expand...
Hopefully Apple can use this app as an internal benchmark to guide them in improving their GPU and Metal backend for Tensorflow and Pytorch.
 
Nermal said:
I just tried this out (a pain to get it all working, but that's beside the point). It seems that the rumours are true: Faces are the stuff of nightmares. Don't try to get it to make people if you want to sleep tonight! :)
Click to expand...

Try increasing the sampling steps. If your installation is defaulting to 25, try 50 or higher. Just keep in mind output time increases.

1661440027115223.jpg
 
Xiao_Xi said:
Has anyone tried running Stable Diffusion with Tensorflow instead of Pytorch? Does it work faster?
github.com

GitHub - divamgupta/stable-diffusion-tensorflow: Stable Diffusion in TensorFlow / Keras

Stable Diffusion in TensorFlow / Keras. Contribute to divamgupta/stable-diffusion-tensorflow development by creating an account on GitHub.
github.com github.com
Click to expand...
I tried Diffusion Bee v0.3.0 (w/ tensorflow backend; M1 Max 10/32/64) and the performance was about the same as v0.1.0. Haven’t had the chance to try the repo you’re specifically asking about, though.
 
altaic said:
I tried Diffusion Bee v0.3.0 (w/ tensorflow backend; M1 Max 10/32/64) and the performance was about the same as v0.1.0.
Click to expand...
That's sad. I would have thought Stable Diffusion could run faster on Tensorflow because Apple has been working on the Metal backend for over a year.

altaic said:
Haven’t had the chance to try the repo you’re specifically asking about, though.
Click to expand...
The dev of Stable Bee and the Stable Diffusion port to Tensorflow is the same, so Stable Bee should use that repo.
 
I've been using InvokeAI, which is a stable-diffusion fork. It can a 50 iteration 512x512 image in 60-80s on my M1 Pro MBP with 16GPU cores. I tried DiffusionBee a couple of weeks back and it was like 5x slower. I just tried DiffusionBee again now and it seems like it's taken a slight lead.

If you do try DiffusionBee, you should be aware that it plops the 4.5GB model in a hidden folder in your home directory. You'll want to delete that if/when you are done playing with the app.
 
Last edited:
  • Like
Reactions: Michaelgtrusa, Tagbert and diamond.g
eas said:
If you do try DiffusionBee, you should be aware that it plops the 4.5GB model in a hidden folder in your home directory. You'll want to delete that if/when you are done playing with the app.
Click to expand...
Thanks, I was wondering where that extra 4.5GB was coming from.
 
What features does one the one-click installer version not have that the brew version does?

Also 'python3 -V' doesn't give me the version number.
 
  • Like
Reactions: rmadsen3
You must log in or register to reply here.
Register on MacRumors! This sidebar will go away, and you'll see fewer ads.
Top Bottom
Back