3 ms·
Diffusion models seem like they're poised to completely replace GANs. They obviously work super well, and you don't have this super finicky minimax training pr
by oofbey 5y ago
Diffusion models seem like they're poised to completely replace GANs. They obviously work super well, and you don't have this super finicky minimax training problem.
- machinekob 5y agoBiggest 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 5y 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
- SleekEagle 5y agoYeah, I haven't seen any big advancements in GANs in few years. Have I missed anything big or is the research volume trending down on them?
- astrange 5y agoThere's this but I don't know if it's been followed up on. https://www.microsoft.com/en-us/research/publication/manifold-matching-via-deep-metric-learning-for-generative-modeling/ https://www.microsoft.com/en-us/research/publication/manifol...
- SleekEagle 5y agoThank you for that link!
- mdda 5y agoOr the two can be combined : https://nvlabs.github.io/denoising-diffusion-gan/index.html https://nvlabs.github.io/denoising-diffusion-gan/index.html
- oofbey 5y agoSounds like it gets the worst of both worlds? The difficult training of a GAN with the slow runtime of a diffusion model.
- mdda 5y agoCould be... Except their page (should you choose to believe it, of course) specifically addresses the advantages: """ "Advantages over Traditional GANs" : Thus, we observe that our model exhibits _better training stability_ and mode coverage. "Why is Sampling from Denoising Diffusion Models so Slow?" : After training, we generate novel instances by sampling from noise and iteratively denoising it _in a few steps_ using our denoising diffusion GAN generator. """