Would the cheapest M1/M2 MBP be OK to run Stable Diffusion on MacOS...

[GMGN]

Yeah I know SD is compatible with M1/M2 Mac but not sure if the cheapest M1/M2 MBP would be enough to run?

Or Specifically if I wish I could get 50 images with 10m with those promts and setting:

Promts
ethereal mystery portal, seen by wanderer boy in middle of woods, vivid colors, fantasy, trending on artstation, artgerm, cgsociety, greg rutkwoski, alphonse mucha, unreal engine, very smooth, high detail, 4 k, concept art, brush strokes, pixiv art, sharp focus, raging dynamic sky, heavens

Seed
3733741481

Guidance scale
7

Dimensions
1024 × 1024


How many GPU/CPU/RAM or which type of M1/M2 would you recommend?

 

[fakestrawberryflavor]

I had no idea what you were talking about in the slightest, so I just googled. Wow, this is scary awesome. I’m interested.

Also, sorry I’m not more help answering your questions lol

 

[Xiao_Xi]

I know SD is compatible with M1/M2 Mac but not sure if the cheapest M1/M2 MBP would be enough to run?

According to the developers of Stable Diffusion:

Stable Diffusion runs on under 10 GB of VRAM on consumer GPUs, generating images at 512x512 pixels in a few seconds.

Stable Diffusion launch announcement — Stability AI

The model itself builds upon the work of the team at CompVis and Runway in their widely used latent diffusion model combined with insights from the conditional diffusion models by our lead generative AI developer Katherine Crowson, Dall-E 2 by Open AI, Imagen by Google Brain and many others. We are

stability.ai

Stable Diffusion uses Pytorch and Pytorch has an experimental Metal backend, so I don't expect Stable Diffusion to be very fast on Apple hardware.

 

[GMGN]

I had no idea what you were talking about in the slightest, so I just googled. Wow, this is scary awesome. I’m interested.

Also, sorry I’m not more help answering your questions lol

haha thanks for replying buddy...

Yeah M1/M2 MBP seems really nice in the long run for its being friendly to AI.

I'm not an exprt on AI at all but yeah...the wave is coming for sure...

 

[GMGN]

According to the developers of Stable Diffusion:

Stable Diffusion launch announcement — Stability AI

The model itself builds upon the work of the team at CompVis and Runway in their widely used latent diffusion model combined with insights from the conditional diffusion models by our lead generative AI developer Katherine Crowson, Dall-E 2 by Open AI, Imagen by Google Brain and many others. We are

stability.ai

Stable Diffusion uses Pytorch and Pytorch has an experimental Metal backend, so I don't expect Stable Diffusion to be very fast on Apple hardware.

hmmm...that would be a bad news for MAC user still have to buy a 3080i?no...please lol...there are tons of GPUs on MBP!!!!!!

 

[leman]

50 images within 10 minutes seems unlikely. If I remember correctly I’ve seen people with M1 reporting around a minute needed for an image (at higher settings).

 

[GMGN]

50 images within 10 minutes seems unlikely. If I remember correctly I’ve seen people with M1 reporting around a minute needed for an image (at higher settings).

Wow...so much time?

So let's buy 3080i ?

 

[leman]

Wow...so much time?

So let's buy 3080i ?

You are comparing a passively cooled ultrakompakte laptop with a humongous desktop GPU. Of course the 3080 ti is going to be much much faster. If ML performance is your primary concern a Mac is a bad choice anyway.

 

[altaic]

…

 

[DearthnVader]

Yeah I know SD is compatible with M1/M2 Mac but not sure if the cheapest M1/M2 MBP would be enough to run?

Or Specifically if I wish I could get 50 images with 10m with those promts and setting:

Promts
ethereal mystery portal, seen by wanderer boy in middle of woods, vivid colors, fantasy, trending on artstation, artgerm, cgsociety, greg rutkwoski, alphonse mucha, unreal engine, very smooth, high detail, 4 k, concept art, brush strokes, pixiv art, sharp focus, raging dynamic sky, heavens

Seed
3733741481

Guidance scale
7

Dimensions
1024 × 1024


How many GPU/CPU/RAM or which type of M1/M2 would you recommend?

That's really cools stuff, if I knew how to use it I could test in on my 13" M2 MBP w/16GB RAM for you.

 

[leman]

To those interested you can try this tool that claims to be a one click installer for stable diffusion: https://github.com/divamgupta/diffusionbee-stable-diffusion-ui

Some discussion: https://news.ycombinator.com/item?id=32804695

 

[DearthnVader]

Yeah I know SD is compatible with M1/M2 Mac but not sure if the cheapest M1/M2 MBP would be enough to run?

Not going to happen with the cheapest M1/M2, you are going to need all the RAM you can get.

Or Specifically if I wish I could get 50 images with 10m with those promts and setting:

Promts
ethereal mystery portal, seen by wanderer boy in middle of woods, vivid colors, fantasy, trending on artstation, artgerm, cgsociety, greg rutkwoski, alphonse mucha, unreal engine, very smooth, high detail, 4 k, concept art, brush strokes, pixiv art, sharp focus, raging dynamic sky, heavens

Seed
3733741481

Guidance scale
7

Dimensions
1024 × 1024


How many GPU/CPU/RAM or which type of M1/M2 would you recommend?

I ran your Promts with Dimensions 768x768:Guidance scale 7:, but I can't set the Seed with the GUI or a the Dimensions higher.

It's using over 35GB of RAM and it's not even started generating anything, the progress bar is not moving and my MacBook Pro can't handle the RAM pressure of all the swap with only 16GB physical memory.

It's not hitting the CPU that hard, but you are going to need tons of RAM to do what you want.

 

[GMGN]

Not going to happen with the cheapest M1/M2, you are going to need all the RAM you can get.

I ran your Promts with Dimensions 768x768:Guidance scale 7:, but I can't set the Seed with the GUI or a the Dimensions higher.

It's using over 35GB of RAM and it's not even started generating anything, the progress bar is not moving and my MacBook Pro can't handle the RAM pressure of all the swap with only 16GB physical memory.

It's not hitting the CPU that hard, but you are going to need tons of RAM to do what you want.

thanks buddy...i think i'll have to use cloud service or buy a 3080i if possible lol

 

[DearthnVader]

thanks buddy...i think i'll have to use cloud service or buy a 3080i if possible lol

Yea, you'd need at least 64GB of RAM to do what you want to do so the Limits you as far as Apple SoC to the M1 Max or the M1 Ultra, and I don't think the GPU can really do it that well in Metal on the M series Apple SoC.

Maybe if it could leverage the Neural Engine and Apple's AI Kit?

This is a really cool program, thank you for bringing it to our attention.

<--I did my profile picture with it real quick!

 

[GMGN]

Yea, you'd need at least 64GB of RAM to do what you want to do so the Limits you as far as Apple SoC to the M1 Max or the M1 Ultra, and I don't think the GPU can really do it that well in Metal on the M series Apple SoC.

Maybe if it could leverage the Neural Engine and Apple's AI Kit?

This is a really cool program, thank you for bringing it to our attention.

<--I did my profile picture with it real quick!

oh buddy your profile picture is maaaaaaaaaaaaaaaaagnificent!!!!

The thing is that i just want to carry one MBP while travelling.

So i'd wait for M3 and see if anything changes...during waiting i'd use cloud GPU service...

Thank you again!

Hurry Up Apple do your own GPU eco-system lol!

 

[nudoru]

Using DiffusionBee I ran your prompt w/ 768x768, 25 steps, 7 gs, and no set seed. On my M1 Max w/ 32gb ram it took 1:56 to generate.

 

[GMGN]

Using DiffusionBee I ran your prompt w/ 768x768, 25 steps, 7 gs, and no set seed. On my M1 Max w/ 32gb ram it took 1:56 to generate.

1m56s or 1h56m???

 

[altaic]

1m56s or 1h56m???

I’ve been messing with DiffusionBee on my MBP M1 Max 64GB and everything completes within a minute or two. Not long to wait for some ****ing weird entertainment.

 

[GMGN]

I’ve been messing with DiffusionBee on my MBP M1 Max 64GB and everything completes within a minute or two. Not long to wait for some ****ing weird entertainment.

1m56s for how many pictures?just one?

 

[wyrdness]

Don't bother with trying to run Stable Diffusion on M1. My M1 Air really struggles with it. Run it in the cloud instead. I'm using replicate.com which provides Nvidia A100 GPU's and is extremely cheap (the free trial is quite generous too). Standard 512x512 images take about 5 seconds.

 

[wyrdness]

This took just over 2 seconds to generate using Stable Diffusion and cost nothing.

Running predict()...
Using seed: 49121

0it [00:00, ?it/s]
2it [00:00, 12.65it/s]
4it [00:00, 12.86it/s]
6it [00:00, 13.21it/s]
8it [00:00, 13.45it/s]
10it [00:00, 13.59it/s]
12it [00:00, 13.67it/s]
14it [00:01, 13.72it/s]
16it [00:01, 13.71it/s]
18it [00:01, 13.51it/s]
20it [00:01, 13.57it/s]
22it [00:01, 13.62it/s]
24it [00:01, 13.65it/s]
26it [00:01, 13.70it/s]
28it [00:02, 13.74it/s]
30it [00:02, 13.79it/s]
30it [00:02, 13.60it/s]

If anyone wants to try it on M1/M2, the inputs were:
width: 512
height: 512
prompt: red-haired woman in enchanted forest in pre-Raphaelite style
guidance_scale: 7.5
prompt_strength: 0.8
num_inference_steps: 30

 

[wyrdness]

Promts
ethereal mystery portal, seen by wanderer boy in middle of woods, vivid colors, fantasy, trending on artstation, artgerm, cgsociety, greg rutkwoski, alphonse mucha, unreal engine, very smooth, high detail, 4 k, concept art, brush strokes, pixiv art, sharp focus, raging dynamic sky, heavens

Seed
3733741481

Guidance scale
7

Dimensions
1024 × 1024

For this image, even an A100 GPU couldn't render it at 1024x1024, because that requires huge amounts of memory. I managed it at 768x768 which took just under 6 seconds.

Here's the price of an A100

 

[GMGN]

For this image, even an A100 GPU couldn't render it at 1024x1024, because that requires huge amounts of memory. I managed it at 768x768 which took just under 6 seconds.

Here's the price of an A100

sounds nice,thanks i'll dive into it!

 

[GMGN]

ok i'll stick into cloud service lol!

 

[altaic]

This took just over 2 seconds to generate using Stable Diffusion and cost nothing.

Running predict()...
Using seed: 49121

0it [00:00, ?it/s]
2it [00:00, 12.65it/s]
4it [00:00, 12.86it/s]
6it [00:00, 13.21it/s]
8it [00:00, 13.45it/s]
10it [00:00, 13.59it/s]
12it [00:00, 13.67it/s]
14it [00:01, 13.72it/s]
16it [00:01, 13.71it/s]
18it [00:01, 13.51it/s]
20it [00:01, 13.57it/s]
22it [00:01, 13.62it/s]
24it [00:01, 13.65it/s]
26it [00:01, 13.70it/s]
28it [00:02, 13.74it/s]
30it [00:02, 13.79it/s]
30it [00:02, 13.60it/s]

If anyone wants to try it on M1/M2, the inputs were:
width: 512
height: 512
prompt: red-haired woman in enchanted forest in pre-Raphaelite style
guidance_scale: 7.5
prompt_strength: 0.8
num_inference_steps: 30

Took 21 seconds with a peak of 14.2 GB RAM utilization and a constant 100% GPU usage on my MBP M1 Max 64GB. Using DiffusionBee, so prompt_strength isn't settable, but all the other settings were as you described. Also, I had a dozen apps open with a couple hundred windows and over a thousand tabs in Safari, so not exactly a best-case benchmarking scenario.