Best ways to run Stable Diffusion

[danqi]

I've been using StableBee and it works quite well, but is very limited.

All other options seem to be quite complicated to set up and much of the information from just a few weeks/months ago is already outdated. What is the most elegant solution right now?

 

[Algr]

I'm interested in this too. What are StableBee's limitations?

I'd like something that can start with my existing sketches and improve and color them while maintaining something of the original. I'm not really a fan of how AI art actually looks at the moment.

 

[danqi]

I'm interested in this too. What are StableBee's limitations?

I'd like something that can start with my existing sketches and improve and color them while maintaining something of the original. I'm not really a fan of how AI art actually looks at the moment.

StableBee is very barebones, lacking most of the options of other builds. But I am pretty new to this as well. But especially the img2img, which is what you would be using, has almost not options for fine-tuning in StableBee.

 

[danqi]

I almost went with trying to install AUTOMATIC1111 but apparently it's sort of semi-supported on Macs and they also had problems with malicious code. Thinking of trying InvokeAI. Has anyone tried that yet?

 

[mi7chy]

InvokeAI had more options than Diffusion Bee when I ran it last month in Sept. It only needed completion time but there have been three recent updates since. I'm back on Big Sur otherwise I would test the updates to see what's new.

 

[Algr]

Those are quite nice. I don't know what most of those settings mean though. You are rendering in 512x512? That seems alarmingly low rez. Do you upscale later or something?

 

[danqi]

Those are quite nice. I don't know what most of those settings mean though. You are rendering in 512x512? That seems alarmingly low rez. Do you upscale later or something?

Yes, the AI is trained on 512x512 images and that is its native resolution.

 

[Algr]

Yikes. I do most work in 4k.

 

[Adreitz]

After generating at 512x512, many people either simply AI upscale or do a cycle of breaking the image up into sections, AI upscaling, running each upscaled section through SD using img2img to regenerate small details, and stitching them back together. InvokeAI automates both options. It's a bit kludgy, but the image can't get any more artificial than it already is.

 

[Algr]

Well I downloaded Diffusion Bee. It seems to run correctly, no crashes. But it is very hard to control output. I was trying to make a Tardis console room, and nothing I can do can produce walls that are not 90°, nor an open uncluttered environment with a table in the middle.
I'm doubtful about this taking work from artists - if you want random pics it is fine, but working towards a novel output seems impossible.

 

[Adreitz]

Note that Apple just released some code enabling Stable Diffusion to run through CoreML. The model files need to be converted for this to work. Apple provides some benchmark results on the GitHub page, but unfortunately they don't compare it to PyTorch SD on the same machines.

GitHub - apple/ml-stable-diffusion: Stable Diffusion with Core ML on Apple Silicon

Stable Diffusion with Core ML on Apple Silicon. Contribute to apple/ml-stable-diffusion development by creating an account on GitHub.

github.com

 

[Algr]

Thanks, but... wow. I don't know where to start with understanding all that.