Most optimized tools for local AI image generation on Macs?

[asiga]

From what I'm finding, in the field of LLMs it seems that llama.cpp is well optimized for current Macs and it even considers them first-class citizens. The same cannot be said for image generation tools, like the popular Comfy UI, where the Mac version is not optimized at all (yes, NVIDIA cards are power-monsters and blah, blah, blah, but the thing is that this guy at reddit was able to substantially improve the Mac performance of Comfy UI by porting DiffusionKit to it, which shows that even if NVIDIA cards are more powerful, the truth is that Comfy UI developers don't optimize for the Mac, and performance is much lower than it could be).

So, I'm wondering: is there any remarkable attempt at getting great AI image performance on current Macs? Everything is changing so quick, that I wouldn't like to be missing something important.

Or at least, any attempt where the Mac is a first-class citizen? (like in the case of llama.cpp for LLMs).

Thanks!

 

[buggz]

Thanks for the reddit link.
While I already have ComyUI installed, I will see if I can update to his diffusion things.
I would like to try FLUX also.
Heh, I am really "winging" this stuff, I am just following the directions online, and it just simply works.

 

[LarTeROn]

Is this a common occurrance? I'm thinking of buying a Mac for AI, but not if it's both overpriced on RAM and unoptimised

 

[bogdanw]

I don’t understand the question. There are many apps designed for Apple Silicon. Are they all that bad?

Mochi Diffusion “Run Stable Diffusion on Mac natively”
https://github.com/MochiDiffusion/MochiDiffusion
Amazing AI “The app is highly optimized and runs on the Apple Neural Engine.”
https://sindresorhus.com/amazing-ai

 

[asiga]

I don’t understand the question. There are many apps designed for Apple Silicon. Are they all that bad?

Mochi Diffusion “Run Stable Diffusion on Mac natively”
https://github.com/MochiDiffusion/MochiDiffusion
Amazing AI “The app is highly optimized and runs on the Apple Neural Engine.”
https://sindresorhus.com/amazing-ai

The question is that Comfy UI seems to be the "de facto standard" for everybody running image/video AI generation locally, and if you perform a simple search, you'll see that the consensus even across Mac users is that not even Max and Ultras can compete with PCs in terms of performance.

I had never heard about Mochi Diffusion, and it's exactly what I was asking for: generative AI tools that run locally on Macs and are well optimized for Apple Silicon.

Yesterday I was looking for video generation on Comfy UI, and it seems that even if it runs on Mac, it's still too slow.

 

[JSRinUK]

Before I came across ComfyUI, I was just using AUTOMATIC1111 (https://stable-diffusion-art.com/automatic1111/) (due to a YouTube vid that showed how to have Ollama generate images from chats). I’m still more comfortable with A1111 than I am with ComfyUI but, as I explore AI video generation, I’m trying to use it more and learn more about it. I’ve been surprised at how little it uses the GPU.

As you can tell, I’m a complete newbie at this.

 

[bogdanw]

DiffusionBee "is the fastest and easiest toolbox to run AI apps locally with Stable Diffusion"
https://diffusionbee.com
"On 8GB M1 MacBook Air, DiffusionBee takes around 30 seconds to generate an image. The speeds are much higher on computers with higher specs."