Does anyone know how hard is it to deploy something like that on CPU instead. While they talk about putting it into consumer hands and the model working on GPU with 10gb vram in seconds, this is still not your average pc. But can that translate to <10min on a CPU instead?
It’s hard to say. Sometimes things are 100x slower on CPU, which is usually tolerable. Sometimes they’re thousands of times slower. Also, remember on Mac M1s there’s an AI accelerator which can speed up TensorFlow if it’s configured correctly even without using the GPU.
iOS devices have “neural engine” hardware too, and I’m seeing equivalents(?) show up in a lot of embedded SoCs.
I’ve been playing with Stable Diffusion quite a bit. I would say it’s not quite as good as DALL-E, but it’s close, and it’s a far smaller model which means it will be practical to run at home (or get more, faster results on their servers). I really hope they release the checkpoints to the public as planned, but I’m afraid they might get cold feet if it impacts their revenue model.