Simply posted a YT-video, evaluating the efficiency of Secure Diffusion Automatic1111 on a Mac M1, a PC with an NVIDIA RTX4090, one other one with a RTX3060 and Google Colab. I personal these machines, so I may give you an perception into my private experiences, benchmarks, pricing and extra. Hope it is useful!
Simply posted a YT-video, evaluating the efficiency of Secure Diffusion Automatic1111 on a Mac M1, a PC with an NVIDIA RTX4090, one other one with a RTX3060 and Google Colab. I personal these machines, so I may give you an perception into my private experiences, benchmarks, pricing and extra. Hope it is useful!
Trying to run A1111 on my M1 Pro, after otherwise being flabbergasted with its performance in traditional applications, was humbling.
Have tried training with the 3060? I am looking at getting 3060 and want to do some simple LoRA training.
For the most part this matches my experiences. My daily driver is also an M1 Macbook Pro Max. I eventually felt like I was using SD enough to get a dedicated midrange PC, so I built my own, although in my case I got the new 4060TI 16GB. It’s definitely faster.
HOWEVER, the difference isn’t as horrific as this test makes it seem. Automatic has made lots of Windows-based optimizations, but I don’t think the Mac has gotten the same amount of attention. A better test IMO would be Draw Things, a dedicated MacOS/iOS app. It’s had several updates focused on optimizing performance since its creation (and anecdotally, those updates have shown).
Repeating your first batch of tests (as well as I could – Draw Things has DPM++ 2M Karras, but not the regular version), I got the following:
– 512×512: 7.3s
– 768×768: 22.3s
– 512×512 + HiRes: 35.9s
– XL 1024: 51.5
On my Windows machine using Invoke (sorry, I have Automatic but generally only use it for specific tasks Invoke can’t handle, so it’s probably not configured very well), I got the following:
– 512×512: 2.5s
– 768×768: 5.2s
– XL 1024: 13.3
I didn’t test Img2Img because I didn’t know your other inputs, but I’d be surprised if the pattern changed.
TLDR: Nothing beats a dedicated graphics card, but the software is just as important.
Linux is probably the fastest.