Been making AI images between my 14 inch MBP and my gaming PC. My findings:

[leman]

If game console makers and Apple can make integrated CPU/GPU chips outperform a dedicated top of the range desktop GPU then why haven't Intel or AMD been able to do what Apple, Sony and Microsoft have been able to do with there integrated CPU/GPU chips.

As others have already said, the PC market is traditionally based on a standardized architecture using physically separate CPU, GPU as well as relatively slow system RAM. The RAM is actually a big problem. Fast GPUs require a lot of memory bandwidth. So a fast integrated CPU+GPU system would need some sort of fast system memory solution. Standard DIMMs become not practical as the cost and power consumption rises quickly and custom solutions are expensive and non-extendable. Apple is pretty much the only company who can afford it since they make the chip for internal consumption.

But it seems that the rest of the industry is going that direction as well, at least for laptops where it makes the most sense. It’s not a coincidence that Intel has been working hard on their own GPU tech. I think it’s just a matter of time before they release some fast CPU+GPU+HBM solution for premium laptops.

 

[Adreitz]

So anyone who has been doing AI image generating on their Macs as well, what prompts have you been using? I mean you already know what I've been doing. =w=

My 2011 Mac Pro with a GTX 970 isn't really capable of running SD and, having a family, I rarely have time to sit down with it.

I was part of the Discord beta using my phone, though, and started out making images based on dreams I had as a kid. SD isn't very good at producing a particular image you have in mind from text, however, so, while the images were generally decent-looking they mostly weren't what I was aiming for. (E.g. the bottom half of a cat, inverted and slithering on the floor with four legs and tail sticking up stiffly in the air; or a miniature volcano erupting in a residential yard; or a room ornately carved out of aquamarine, dimly lit and with sunlight filtering through the walls.)

I then started making less specific images based on random ideas: a faceless man in a tuxedo sitting in a bonfire (SD refused to understand the concept of a person without a face), a baby drinking whisky, children playing happily in a yard while their house burns behind them, etc.

I'm thinking about getting an M2 Max MBP when they're released so I can try running SD locally.

 

[Spaceboi Scaphandre]

My 2011 Mac Pro with a GTX 970 isn't really capable of running SD and, having a family, I rarely have time to sit down with it.

I was part of the Discord beta using my phone, though, and started out making images based on dreams I had as a kid. SD isn't very good at producing a particular image you have in mind from text, however, so, while the images were generally decent-looking they mostly weren't what I was aiming for. (E.g. the bottom half of a cat, inverted and slithering on the floor with four legs and tail sticking up stiffly in the air; or a miniature volcano erupting in a residential yard; or a room ornately carved out of aquamarine, dimly lit and with sunlight filtering through the walls.)

I then started making less specific images based on random ideas: a faceless man in a tuxedo sitting in a bonfire (SD refused to understand the concept of a person without a face), a baby drinking whisky, children playing happily in a yard while their house burns behind them, etc.

I'm thinking about getting an M2 Max MBP when they're released so I can try running SD locally.

Damn you're still using a 2011 Mac Pro? 👀

Gonna say, The Max chip unless you do graphical work or video editing you don't need it. The M1 Pro is more than enough. Hell it's still more than what I use a computer for and I'm a PC gamer. If you're getting a MBP M2 Pro will handle it nicely.

Though if you want M2 Max the Mac Studio's probably gonna be an option. I'd recommend the Max more in that.

 

[Adreitz]

Damn you're still using a 2011 Mac Pro? 👀

Gonna say, The Max chip unless you do graphical work or video editing you don't need it. The M1 Pro is more than enough. Hell it's still more than what I use a computer for and I'm a PC gamer. If you're getting a MBP M2 Pro will handle it nicely.

Though if you want M2 Max the Mac Studio's probably gonna be an option. I'd recommend the Max more in that.

On the contrary, video editing will use the hardware codecs, which I believe are the same between the Pro and Max, while more GPU cores will directly translate into improved rendering speed with SD.

 

[MysticCow]

It's not about Apple being out of touch with the gaming industry. It's about Windows is the current standard and game developers have no motivation to develop for a platform that very few gamers use. It's kind of like what came first the chicken or the egg. Perhaps if game developers made more games for macOS then people would play them but people aren't playing them because game developers aren't making them.

As someone from Aspyr once famously said, and I am paraphrasing from MacWorld:

"1 GHz is 1 GHz, regardless of processor architecture. And if you want timely ports from Win to Mac, you aren't getting Altivec support, either."

Just adding here, not criticizing. Please don't hate me.

There are a lot of factors playing into no Mac games. Market share. Apple's general attitude. Different processor architectures now.

Simply put, there's a lot of road to be made up here...

 

[Spaceboi Scaphandre]

On the contrary, video editing will use the hardware codecs, which I believe are the same between the Pro and Max, while more GPU cores will directly translate into improved rendering speed with SD.

Yeah but will you need that power longterm when the AI image generating fad dies out and it goes to just the few who genuinely use it often instead of the people coming to check out the "hot new technology?" I didn't need the Max's longterm power which is why my MBP is a base spec M1 Pro.

 

[Unregistered 4U]

Yeah but will you need that power longterm when the AI image generating fad dies out and it goes to just the few who genuinely use it often instead of the people coming to check out the "hot new technology?" I didn't need the Max's longterm power which is why my MBP is a base spec M1 Pro.

Microsoft is reportedly building that AI image generating fad into the upcoming Office (renamed 365). Need a background for a slide? “rolling hills with dark clouds on the horizon” Need some clip art to drop into that party flyer? “cheerful large anthropomorphic red balloons” and “table festooned with party snacks and drinks” or some such.

I figure it’s going to be one of those things where folks in the future wonder, “So, you just bought thousands and thousands of clip art images without even knowing if any of those images were going to be useful for your particular need? And then, if it wasn’t, you had to kludge/edit several of them (with varying styles) to combine into one?” It’ll just be one of those “table stakes” that everyone will expect computers will be able to do quickly. “cloud of question marks around a detailed plastic figure shrugging it’s shoulders”

 

[Spaceboi Scaphandre]

Microsoft is reportedly building that AI image generating fad into the upcoming Office (renamed 365). Need a background for a slide? “rolling hills with dark clouds on the horizon” Need some clip art to drop into that party flyer? “cheerful large anthropomorphic red balloons” and “table festooned with party snacks and drinks” or some such.

I figure it’s going to be one of those things where folks in the future wonder, “So, you just bought thousands and thousands of clip art images without even knowing if any of those images were going to be useful for your particular need? And then, if it wasn’t, you had to kludge/edit several of them (with varying styles) to combine into one?” It’ll just be one of those “table stakes” that everyone will expect computers will be able to do quickly. “cloud of question marks around a detailed plastic figure shrugging it’s shoulders”

Well there's some uses for the AI generating for making B assets, like background music as Unreal Engine 5 has an AI that can create simple background music as demonstrated in the downtime between Fortnite Chapter 3 Season 2 and 3 as all the music that was used during downtime was made by AI

It can also be used for stock images for presentations. For bigger more detailed stuff, nah human artists still superior.

 

[ondioline]

Here's 10 images at 768x768 80 steps in 8 mins 39 secs on laptop 3060 so 100 images is <= 1 hour 26 mins

Since people aren't posting comparisons heres mine. At 512x768, 50 steps, it takes 43-45 seconds to make 1 image. So about 1 hour and 22 minutes to generate 100.

This is a 16 MBP with M1 Max 32 core GPU. So it's apparently similar to the 3060 mobile.

Now I also tested it on a Radeon 5700XT and it was taking 3 minutes+ per image so, AMD is apparently awful for this.

 

[ondioline]

Either way, I think the mps backend is inferior to the CUDA backend in pytorch. Apple Silicon is adequate for SD but I wouldn't run out and buy an M1 Ultra for this.

 

[Unregistered 4U]

For bigger more detailed stuff, nah human artists still superior.

Superiority doesn’t matter in a very large number of cases, though, as cheap/free can quickly outweigh superiority.

THIS isn’t even a very good image, but it gets the point across and if the user likes it (and this is just an invite for a block party), then that’s all that’s needed. For every company willing to spend a few thousand dollars on professionally produced invitations, there’s at least a thousand unskilled grade schoolers that will simply be sprucing up their book report, high schoolers adorning their science project, and college students soundtracking their student project that won’t think twice to have AI toss out something that they don’t have to fuss over.

I guess what I’m saying is that, to me, it’s like the first folks that were editing images on their large powerful workstations and musing to themselves “Your average person won’t need this kind of power, though. This is just a fad for us photographers that like to play around with digital editing” and now digital editing of high resolution photos at high bit depths is something everyone’s doing on their phones. Are human artists still better than Photos at balancing an image? Very likely. But, there are millions of folks every day just clicking on the magic wand in Photos and being pleased with the results.

 

[Adreitz]

Now I also tested it on a Radeon 5700XT and it was taking 3 minutes+ per image so, AMD is apparently awful for this.

Thanks for posting your numbers. On the PC side, I believe SD is reliant on CUDA by default, which is nVidia-specific. Your 5700 machine is probably falling back to the CPU. I've seen people asking about the status of AMD compatibility, but I don't know whether or not anyone has gotten it working in any repo.

[Edit] Maybe check out this guide: https://www.travelneil.com/stable-diffusion-windows-amd.html [/Edit]

 

[Adreitz]

Either way, I think the mps backend is inferior to the CUDA backend in pytorch. Apple Silicon is adequate for SD but I wouldn't run out and buy an M1 Ultra for this.

I think MPS support is very new in PyTorch. A lot of operations are still not supported. So things will hopefully improve over time.

[Edit] The PyTorch operation support status is here: https://github.com/pytorch/pytorch/issues/77764 [/Edit]

 

[russell_314]

As someone from Aspyr once famously said, and I am paraphrasing from MacWorld:

"1 GHz is 1 GHz, regardless of processor architecture. And if you want timely ports from Win to Mac, you aren't getting Altivec support, either."

Just adding here, not criticizing. Please don't hate me.

There are a lot of factors playing into no Mac games. Market share. Apple's general attitude. Different processor architectures now.

Simply put, there's a lot of road to be made up here...

I agree there’s a lot of factors. I’d say a major one is Apple’s decision to not pursue it. This would be an expensive pursuit, and I have a feeling Apple feels not a very profitable one.

 

[jerryk]

So I found what I did wrong on PC. CUDA was not installed. It's now rendering faster than my Mac. However it's still loud, and my images on PC are making some really weird mutations. I find that the renders I do on Mac are a lot more stable. There's some renders the Mac does faster on too such as k_lms.

So I guess it goes back to that Mac Address video. The x86 PCs are faster but jankier while the ARM Macs are a lot more precise and stable.

But, are you running the same code base on both machines? I would assume the changes in code make more of a difference than the graphics processor or OS that the code runs on.

 

[Xiao_Xi]

Stable diffusion runs faster on macOS using Tensorflow than using Pytorch.

Speed Up Stable Diffusion on Your M1Pro Macbook Pro

How to speed up your Stable Diffusion inference and get it running as fast as possible on your M1Pro Macbook Pro laptop. Made by Thomas Capelle using W&B

wandb.ai

I've seen people asking about the status of AMD compatibility, but I don't know whether or not anyone has gotten it working in any repo.

Pytorch supports AMD GPUs on Linux only.

 

[Spaceboi Scaphandre]

Stable diffusion runs faster on macOS using Tensorflow than using Pytorch.

Speed Up Stable Diffusion on Your M1Pro Macbook Pro

How to speed up your Stable Diffusion inference and get it running as fast as possible on your M1Pro Macbook Pro laptop. Made by Thomas Capelle using W&B

wandb.ai

Pytorch supports AMD GPUs on Linux only.

This gives me an idea for using the AIs on a Steam Deck then...

 

[ondioline]

Pytorch supports AMD GPUs on Linux only.

On the PC side, I believe SD is reliant on CUDA by default, which is nVidia-specific. Your 5700 machine is probably falling back to the CPU.

You can use mps with a Radeon, this was on MacOS.

 

[leman]

Either way, I think the mps backend is inferior to the CUDA backend in pytorch. Apple Silicon is adequate for SD but I wouldn't run out and buy an M1 Ultra for this.

Oh, there is no question about that. CUDA backend is not only much more mature but can also harness the tensor cores in the Nvidia GPUs. The MPS backend instead relies on regular compute shaders which will be much slower. I wonder how Apple intends to proceed here. They are continuously improving both the AMX engine and the NPU, but those are not exposed to the end user which makes things difficult. Maybe in the future there will be some sort of hardware and API unification, with Metal becoming the API for the GPU and the tensor engine…

 

[FSMBP]

This gives me an idea for using the AIs on a Steam Deck then...

Very awesome thread and good insight! Hoping DiffusionBee keeps updating!

 

[Spaceboi Scaphandre]

Very awesome thread and good insight! Hoping DiffusionBee keeps updating!

I probably won't do any rendering on my Deck though as my storage is pretty full, plus my Deck is for games and games only. I'm not one of those few who actually use the Steam Deck as a workstation (though it can be used as one)

 

[mi7chy]

Since people aren't posting comparisons heres mine. At 512x768, 50 steps, it takes 43-45 seconds to make 1 image. So about 1 hour and 22 minutes to generate 100.

This is a 16 MBP with M1 Max 32 core GPU. So it's apparently similar to the 3060 mobile.

Now I also tested it on a Radeon 5700XT and it was taking 3 minutes+ per image so, AMD is apparently awful for this.

Retested for 512x768 50 steps which took 16.87s.

Also have Staple Diffusion running on RX6800 but software stack is keeping it back since other GPU compute workloads like hashcat compete closely with equivalent Nvidia desktop GPU.

 

[Adreitz]

Stable diffusion runs faster on macOS using Tensorflow than using Pytorch.

Speed Up Stable Diffusion on Your M1Pro Macbook Pro

How to speed up your Stable Diffusion inference and get it running as fast as possible on your M1Pro Macbook Pro laptop. Made by Thomas Capelle using W&B

wandb.ai

Thanks! I don't really see much in the way of a how-to here, though. They do a performance comparison, but I don't understand how they got it running. And I can't tell how a current Mac-compatible repo (such as InvokeAI) compares with either of these results. I get the feeling that W&B is a very insular group and assumes a high level of understanding of ML and coding in general.

 

[jerryk]

Microsoft is reportedly building that AI image generating fad into the upcoming Office (renamed 365). Need a background for a slide? “rolling hills with dark clouds on the horizon” Need some clip art to drop into that party flyer? “cheerful large anthropomorphic red balloons” and “table festooned with party snacks and drinks” or some such.

I figure it’s going to be one of those things where folks in the future wonder, “So, you just bought thousands and thousands of clip art images without even knowing if any of those images were going to be useful for your particular need? And then, if it wasn’t, you had to kludge/edit several of them (with varying styles) to combine into one?” It’ll just be one of those “table stakes” that everyone will expect computers will be able to do quickly. “cloud of question marks around a detailed plastic figure shrugging it’s shoulders”

Office 365 Professional has had a lookup of millions of sourced images for years. In PowerPoint, and possible elsewhere, you can enter the prompts like you listed and get very high-quality images to use in your slides. Currently, these are actually images, not generated, that Microsoft has licensed for you to use. They probably use some sort of Text encoded model to power the search similar to what these Art generators are doing. I suspect they will add some novel image generator capability at some time. The only catch is I believe you still have to have a paid Office/Microsoft 365 subscription to get this feature. Someone has to pay for all those compute cycles.

 

[Xiao_Xi]

I don't really see much in the way of a how-to here, though. They do a performance comparison, but I don't understand how they got it running.

DiffusionBee uses Tensorflow-based Stable Diffusion since 0.2.

• Tensorflow backend for 4x speed on 8GB Macs

Release 0.2.0 · divamgupta/diffusionbee-stable-diffusion-ui

New features: Stop Button Multiple Images Tensorflow backend for 4x speed on 8GB Macs Old OS versions also supported now History Styles

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