3 ms·
Biggest problem for diffusion models were performance (as you need to iterate even at inference) But I'm not up to date with newest architectures maybe its alre
by machinekob 4y ago
Biggest problem for diffusion models were performance (as you need to iterate even at inference) But I'm not up to date with newest architectures maybe its already solved :P
- johndough 4y agoI was wondering it if would be possible to train a neural network to do multiple iterative steps at once. As it turns out, it has already been done and it requires about 4 to 8 distilled iterations for comparable quality. If this pace keeps up, we will probably see similar running time to GANs in the near future. https://arxiv.org/pdf/2202.00512.pdf https://arxiv.org/pdf/2202.00512.pdf