Tackling the Generative Learning Trilemma with Denoising Diffusion GANs
I prepared a presentation about this ICLR 2022 Spotlight paper for METU CENG 796 Deep Generative Models course in Spring 2023. Since I noticed that there are almost no videos about this paper on the internet, I decided to record it and publish it.
Here’s the video:
I don’t like the white background in the presentation, but I didn’t bother to change it since all figures from the paper had white backgrounds.
Resources & References
- Presentation slides
- Zhisheng Xiao, Karsten Kreis, Arash Vahdat (2021). Tackling the Generative Learning Trilemma with Denoising Diffusion GANs.
- Project page of the paper
- GitHub repository of the paper
- Jonathan Ho, Ajay Jain, Pieter Abbeel (2020). Denoising Diffusion Probabilistic Models.
- Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, Yoshua Bengio (2014). Generative Adversarial Networks.
Comments
The paper and the reported results are quite impressive. However, to the best of my knowledge, there are no real-life applications that use the method proposed in this paper. Diffusion models was booming in the last year; we have seen great applications such as Stable Diffusion, Midjourney, and so on. Despite the reported benefits of this method, none of the diffusion applications used this method.
The GitHub repo of the paper is basically abandoned: only 3 commits, and the latest commit dates back to the last year. It’s a shame that they didn’t improve it further. One of the possible reasons, as hypothesized by our course instructor Gökberk Cinbiş, is that people didn’t want to work on this line of research because GANs are notoriously hard to train.