swift-diffusion VS ncnn

I am going to put model related code we use in a public repo soon (it is very similar to https://github.com/liuliu/swift-diffusion but in NHWC format). ANE will be around 25s if it runs. DT's default only uses GPUs and 35s is on GPU (yes, like you said, upscaling would take extra 10s).

The text encoder uses a new vocabulary set, make sure you copied them from open_clip repo: https://github.com/mlfoundations/open_clip (I have these also available at: https://github.com/liuliu/swift-diffusion/tree/liu/unet/examples/open_clip

Should be able too, if there is a need. I am more interested to support training hypernetwork from the app directly. The conversion script itself is open-source (https://github.com/liuliu/swift-diffusion/blob/main/examples/unet/main.swift), but not polished, and because Apple doesn't allow you to run Python on device, so I cannot make it as easy as typing a URL and get done. Need to figure out what the UX looks like without me providing a networked services ...

Hi, this is the first app in a while (probably 10 years) that I submitted to AppStore. I've done this app in 3 weeks, so there are a lot to be polished. The technology that enables this I discussed in depth in an accompanied blog post: https://liuliu.me/eyes/stretch-iphone-to-its-limit-a-2gib-mo...

Some parts of it (or major parts) is also available at https://github.com/liuliu/swift-diffusion. I plan to port more stuff back to swift-diffusion and make a CLI tool out of it (it is a bit more work than the app because I need to consider CUDA compatibility there).

AMA!

ncnn uses Vulkan for GPU acceleration, I've seen it used in a few projects to get AMD hardware support.

https://github.com/Tencent/ncnn

They're using NCNN to package the model. Have a look. https://github.com/Tencent/NCNN

Hi. Here is my prefered android app for realtime objet detection: https://github.com/nihui/ncnn-android-nanodet ; https://github.com/Tencent/ncnn contains a lot of android demo app for a lot of models.

ncnn home page: https://github.com/Tencent/ncnn

ncnn

Yes, Android devices tend to have bigger RAMs, making running 1024x1024 possible (this is not possible at all on iPhones, which could peak around 5GiB memory with my current implementation, some serious engineering required to bring that down on iPhone devices). The problem is I am not sure about speed. I would likely switch to NCNN (https://github.com/Tencent/ncnn) as the backend which have a decent Vulkan computing kernel support. It is definitely a possibility and there is a path to do that.

As the leading open-source inference framework in China and in the world, what we like are its almost zero cost cross-platform capability, high inference speed, and minimal deployment volume. (Project address: https://github.com/Tencent/ncnn)

If you’re not planning on training the neural net on an embedded device and just do inference, this might interest you: https://github.com/Tencent/ncnn