A.I. help!!! Image creation.

[AlienMonopoly]

I need help with training diffusion.

 

I am so far removed from my comfort zone.

 

I installed hombrew from my m4 Mac mini using terminal. WTF!!!
 

1st step is admitting you need help. HELP lol.

 

I thought I could just use LM studio. But none of the models would create an image.

 

So then I installed hombrew (might want to uninstall) and installed diffusion and tried it using the webui. 
 

It made an image out of a horror movie.

 

i looked and saw that I could train it. Well how? 

[Eigenvektor]

6 hours ago, AlienMonopoly said:

i looked and saw that I could train it. Well how? 

What you need is:

• A GPU (preferably multiple), as powerful as possible
• A lot of data to train on
• The time to train the model

The full process is much too involved for a forum. You may want to Google for tutorials such as these:

• https://medium.com/@geronimo7/training-a-latent-diffusion-model-from-scratch-897c7b77ece9
• https://www.hyperstack.cloud/technical-resources/tutorials/how-to-train-a-stable-diffusion-model

[05032-Mendicant-Bias]

Apple is bad for diffusion. You can do LLM on Apple just fine because they are easier to accelerate even Vulkan will work great, but for diffusion you need a properly implemented ML framework, you need something that accelerates pytorch, and that is ONLY possible with Nvidia CUDA. Nobody else in the world managed to do a easy to use and properly accelerated pytorch binary binding. Nvidia cards cost twice or thrice and still are worth the premium if you want to do ML.

 

I'm using AMD ROCm and it took me six months to get Comfy UI to accelerate with Pytorch. It's hardcore. With Apple Metal there is no hope, unless you want to write pytorch binaries yourself.

 

That said you can run diffusion on Apple, just with CPU acceleration on those poor weak firestorm and icestorm cores.

 

Use SD Next and it'll work with a one click installer. It's just you take five minutes to do what my AMD card does in twenty seconds. Tamper your expectation.

 

https://github.com/vladmandic/sdnext

4 hours ago, AlienMonopoly said:

I need help with training diffusion.

You do not. You aren't going to do that on an apple, ever.

 

Learn SD Next, and learn img2img and use LoRA that other people trained. You don't need a whole new diffusion model to do what you want to do. Just learn what's already there.