Stable Diffusion image size on m1 Max 64gb ram

[Mac3Duser]

Hi,
I would like to know if someone has already tested stable diffusion in command line on an apple M1 max and what is the larger image size possible.
As the unified memory is also the vram, and as image size depends on the vram, if the m1 max has 64gb ram, is it possible to make natively images bigger than 1024 x 1024 without upscaling ?

 

[l0stl0rd]

Probably yes but the last version I tried on Mac still seems slow and buggy.

Strangely I could not with my 32 GB M1 Pro.

I will check later but 768x1024 definitely works on my 3080 but not on my M1 Pro with 32 GB 🤷‍♂️

After finally getting Automatic 1111 to for on Mac OS.

Update:

so yes the M1 Pro 32 GB can do 1024x1024 but it is very slow, like 2 min for 20 sampling steps with Euler a
As a comparison my 3080 can do 2048x2048 in about the same time.

768x768 is a lot more usable on my Mac at like 45 sec.

Can’t test it with Diffusion Bee as it has a max of 768x768.

Hope this helps.

 

[Mac3Duser]

thanks a lot for your answer.
with my 3090 (24 go vram) I can't do 2048x2048, so maybe you are upscaling with real-ERSgan ? the limit is something like 1024 x 896
interesting if your M1 pro 32gb can do 1024x1024, have you any indications of the amount of Vram used in the process ?

 

[l0stl0rd]

thanks a lot for your answer.
with my 3090 (24 go vram) I can't do 2048x2048, so maybe you are upscaling with real-ERSgan ? the limit is something like 1024 x 896
interesting if your M1 pro 32gb can do 1024x1024, have you any indications of the amount of Vram used in the process ?

No the 3080 is not upscaling but I noticed too that Vram use of Automatic 1111 has gotten better. Just for info my 3080 has 12 GB not 8.

Also this is with a fresh start so I would not bet on it still working after a while but not sure.

Not sure about the Vram use as there in some ways there is none and memory load I can check but is was fairly full like 70-80 % I would guess from iStats that I have running.

 

[Mac3Duser]

ok so you use the command --lowvram ?

 

[l0stl0rd]

ok so you use the command --lowvram ?

nothing I just run the webui.bat or sh without adding anything. It is however a very recent version as I wanted to try the SD 2.0 model.

 

[h9826790]

That should work. Even my 8GB base model M1 Macbook Air can do 768x768 flawlessly. I can't see why a 64GB model cannot do 1024x1024 and above.

For info, I never try 1024 on my M1 Macbook Air because I use DiffusionBee, and no such option exist in the GUI.

Anyway, I've tested higher resolution image generation on my Mac Pro with Radeon VII (16GB VRAM), and a PC with 1080Ti (11GB VRAM). Both of them can generate pictures about 1024x1024. But from my own observation, this doesn't work well most of the time. I guess becuase the source pictures for training aren't at that high resolution. Therefore, when we ask for a high resolution picture, the software just combine few lower resolution pictures into a bigger one.

e.g. If I want to generate a picture of a person. 768x768 or lower usually have good outcome.

But once I move to 1536x1536, then I got four people on a single picture (for the same seed number). It feels like the software just combine four 768x768 source picture into one.

No the 3080 is not upscaling but I noticed too that Vram use of Automatic 1111 has gotten better. Just for info my 3080 has 12 GB not 8.

Also this is with a fresh start so I would not bet on it still working after a while but not sure.

Not sure about the Vram use as there in some ways there is none and memory load I can check but is was fairly full like 70-80 % I would guess from iStats that I have running.

You may try the terminal command IOAccelMemory. This may allow you to check how much VRAM is occupied by a specific software.

 

[Mac3Duser]

thank you very much, I will try with other sizes, and not square... but with windows I know how to make larger images but not with macs silicon

 

[leman]

Did anyone try the new tools Apple has published? They should be more memory efficient

 

[l0stl0rd]

That should work. Even my 8GB base model M1 Macbook Air can do 768x768 flawlessly. I can't see why a 64GB model cannot do 1024x1024 and above.

For info, I never try 1024 on my M1 Macbook Air because I use DiffusionBee, and no such option exist in the GUI.

Anyway, I've tested higher resolution image generation on my Mac Pro with Radeon VII (16GB VRAM), and a PC with 1080Ti (11GB VRAM). Both of them can generate pictures about 1024x1024. But from my own observation, this doesn't work well most of the time. I guess becuase the source pictures for training aren't at that high resolution. Therefore, when we ask for a high resolution picture, the software just combine few lower resolution pictures into a bigger one.

e.g. If I want to generate a picture of a person. 768x768 or lower usually have good outcome.
View attachment 2126916

But once I move to 1536x1536, then I got four people on a single picture (for the same seed number). It feels like the software just combine four 768x768 source picture into one.
View attachment 2126915


You may try the terminal command IOAccelMemory. This may allow you to check how much VRAM is occupied by a specific software.

Normal most models are optimized for 512x512 px abd SD 2 is 768x768 px once you go a lot bigger it gets tougher to get good results.

 

[Mac3Duser]

it is quite necessary to have high resolution especially for the img2img
and with an RTX 3090 I get interesting images in txt2img 1024 x 640 for example