I think it's unclear how much computational resources the uncompression steps take. At the moment it's fairly fast, but RAM hungry. But this article makes it c

I think it's unclear how much computational resources the uncompression steps take. At the moment it's fairly fast, but RAM hungry. But this article makes it clear that quantizing the representation works well (at least for the VAE). It's possible quantized models could also do decent jobs.