Skip to content
The methodOpen data
Back to the collection

Progressive Growing of GANs for Improved Quality, Stability, and Variation

Tero Karras, Timo Aila, Samuli Laine, Jaakko Lehtinen

2017 · arXiv

Other research
Read the paper
HN score
0
Mentions
4
Readers
4
Months
4
What’s in a score?

From the abstract

We describe a new training methodology for generative adversarial networks. The key idea is to grow both the generator and discriminator progressively: starting from a low resolution, we add new layers that model increasingly fine details as training progresses. This both speeds the training up and greatly stabilizes it, allowing us to produce images of unprecedented quality, e.g., CelebA images at 1024^2. We also propose a simple way to increase the variation in generated images, and achieve a record inception score of 8.80 in unsupervised CIFAR10. Additionally, we describe several implementation details that are important for discouraging unhealthy competition between the generator and discriminator. Finally, we suggest a new metric for evaluating GAN results, both in terms of image quality and variation. As an additional contribution, we construct a higher-quality version of the CelebA dataset.

Metadata: arXiv. Abstracts retain their original authors’ rights.

In the conversation

Original HN comments, with labels from simple text rules. A label describes the whole comment and may miss context or irony. Read the discussion to make up your own mind. About the labels

4 selected excerpts · 4 total mentions

godelski
Neutral mention · rule
> I'm disgusted, but not surprised, that teenage kids are generating CSAM like this. Even before we had diffusion models, we had GANs and deepfakes, which were almost immediately used for generating shittons of nonconsensual porn I think the big difference now is that 1) it's much easier to do now, and 2) the computational requirements and (more importantly) technical skills have dramatically dropped. We should also be explicitly aware that deep fakes are still new. GANs in 2014 were not creating high definition images. They were doing fuzzy black and white 28x28 faces, poorly, and 32x32 color images that if you squint hard enough you could see a dog…
Read the full comment
(https://arxiv.org/abs/1406.2661). MNIST was a hard problem at that time and that's 10 years. It took another 4 years to get realistic faces and objects (https://arxiv.org/abs/1710.10196) (mind you, those images are not random samples), another year to get to high resolution, and another 2 to get to diffusion and another 2 before those exploded. Deep fakes were really only a thing within the last 5 years and certainly not on consumer hardware. I don't think the legal system moves much in 10 years let alone 5 or 2. I think a lot of us have not accurately encoded how quickly this whole space has changed. (image synthesis is my research area btw) I'm not surprised that these teenagers in a small town did this. But the fact that all those adjectives exist in that order is distinct. Discussions of deep fakes like that Tom Scott video were barely a warning (5 years is not a long time). It quickly went from researchers thinking it can happen in the next decade and starting discussions to real world examples making the news in under their prediction time (I don't think anyone expected how much money and man hours would be dumped into AI).
Original comment SDXL Turbo: A Real-Time Text-to-Image Generation Model
NegatioN
Neutral mention · rule
I don't know exactly what was used to upscale here, but Progressive Growing of GANs[0] was a breakthrough last year that proved at least 1024x1024 was viably producable. The short explanation is: Train and freeze the most basic layer of the model progressively to "let the network" understand higher resolution concepts one piece at the time, and avoid mode-collapse. The network architecture illustrating this a bit better is shown on page 3 of the paper [1] [0]: https://github.com/tkarras/progressive_growing_of_gans [1]: https://arxiv.org/pdf/1710.10196.pdf
Original comment Christie’s sells its first AI portrait for $432k
nl
Neutral mention · rule
GANs (which he invented) are a game changer, and have enabled things to be built which I thought were 20 years away. For example, the life-like faces NVidia research created in Progressive Growing of GANs for Improved Quality, Stability, and Variation[1] are just an incredible achievement (it's worth looking at the pics in that paper if you haven't seen them). Approaches like that have real-life impact in things like games, architecture, design, film, etc (and that is just the obvious implications for images). GANs of course can be used outside images and they have just as much potential there. Anywhere an auto-encoder is used now can probably be done better with a GAN. [1] https://arxiv.org/abs/1710.10196
Original comment A.I. Researchers Are Making More Than $1M, Even at a Nonprofit
nealeratzlaff
Neutral mention · rule
I think for most people that hurdle of understanding is immense, and isn't likely to get any better. ML/DL infrastructure will continue to improve and deploying AI systems will become more accessible as time goes on. But I don't see familiarity with (non) convex optimization, statistical learning, topology, etc. becoming mainstream. Without these its hard to see a reasonable path forward for AI. For me this year in DL was a lot narrower. it mostly revolved around systems becoming more human in capability. I think most of this is in the article already: We can now hopefully say that ImageNet has been solved. AlphaZero tackled structured games with a single algorithm…
Read the full comment
[1]. Tacotron2 produced completely passible TTS [2] GANs improved dramatically and saw some good theory to back it up [3] [4]. We also started to care more about how well our models do in general, which to me shows maturity: Adversarial examples showed their teeth, and hopefully convinced everyone to care about robust models [5] [6]. Reinforcement learning algorithms were shown to have poor transferability [7]. In 2018 I hope to see a new kind of CNN. Residual style networks are the norm now, since they mostly solve problems with gradient flow. But take away all the skip connections and we're mostly left with a vgg-style linear net with box filters. I'd be really excited to see a network with image-sized conv filters that could adapt their shape (and therefore representational power) to a given feature or signal. Hopefully 2018 is the year where people stop calling AI a one-trick-pony. I don't hear it as often these days, but I think its time to put that phrase to rest. [1] https://arxiv.org/abs/1712.01815 [2] https://research.googleblog.com/2017/12/tacotron-2-generatin... [3] https://arxiv.org/abs/1710.10196 [4] https://arxiv.org/abs/1701.04862 [5] https://arxiv.org/abs/1707.07397 [6] https://arxiv.org/abs/1710.06081 [7] https://arxiv.org/abs/1709.06560
Original comment AI and Deep Learning in 2017 – A Year in Review

Keep following the thread.

More in Other research

More from the same topics.