Any thoughts on upcoming image generation support?

[DalekoDaleko]

It looks like Image Generation support is coming to Ollama (see https://ollama.com/[x/z-image-turbo](https://ollama.com/x/z-image-turbo)), starting with Mac, then Windows/Linux. Is there something that could be done today to get ready to use this in OllamaSharp? (could it be that we could already start using it although without the support in the Ollama UI client?)

[awaescher]

They do have a very basic api endpoint which is described in their docs. But it's marked as experimental with a disclaimer that the api might change anytime:

Image generation is experimental and may change in future versions.

They also added this:

Image generation is now supported through the standard /api/generate endpoint when using image generation models. The API automatically detects when an image generation model is being used.

So I guess /api/generate could be affected as well, but there's no final spec yet.

[DalekoDaleko]

Great, sounds like we could start drafting the wrapper(s)... Do you accept contributions? :)

[awaescher]

Sure.

[DalekoDaleko]

I am not an AI expert, I am just looking to use local generation in one of my projects, but here is my semi-educated opinion:

I looked a bit closer at what they did for Mac OS and it's all based on Apple MLX platform and library, which I suppose does all the heavy lifting for them. And the only relatively straightforward way to run those models they are experimenting with (like Z-Image) on Windows right now is to use stable diffusion, perhaps through stable-diffusion.cpp. There is also Olive which kind-of supports stable-diffusion models but with lots of limitations and broken examples left as a homework for the readers. So, I am somewhat skeptical that we should expect that Windows/Linux support in Ollama any time soon - it would be a whole separate undertaking, and the APIs would likely have to be refactored to unify across OSes.