256 points by hyunjkliz 5 months ago flag hide 16 comments
deeplearning_fan 5 months ago next
Great article! I've been working with GANs for a while but I haven't seen such realistic images before. Can't wait to try it out on my projects.
researcher_alice 5 months ago next
Thanks for your kind words. It took us months to generate these images and we hope they will inspire more people to explore the limits of deep learning with GANs.
deeplearning_fan 5 months ago next
We used 4 NVIDIA A100 Tensor Core GPUs with 40GB of memory. They offer the highest performance and memory capacity for large-scale generative models. You may want to consider using a cloud-based solution, like Google Colab, to gain access to high-end GPUs without breaking the bank.
user_bob 5 months ago prev next
These results are impressive, but hard to replicate without the code. Any plan to make the code open source?
deeplearning_fan 5 months ago next
Yes, we will release the code in the next few weeks. We plan to open source the full project, including the implementation and dataset, so stay tuned for updates!
startup_john 5 months ago prev next
Is it possible to use these generative models for commercial use? Like creating a seamingless, high-quality avatars for an NFT market place?
researcher_alice 5 months ago next
As long as the use case complies with our open source license, there should be no issue. However, we encourage you to seek legal counsel to ensure full compliance with the law.
ai_engineer_mike 5 months ago prev next
I'm curious, what kind of optimization algorithm are you using for the GANs training process? Our team had some instability issues with our old implementation.
deeplearning_fan 5 months ago next
We used DCGAN with a bit of WGAN to improve the stability. It took a lot of experimentation to find the best setup, but we also had some luck by using the zero-centered gradient penalty method. Hope it helps!
hacker_sam 5 months ago prev next
Not sure if this is helpful, but I used some of the popular generative models from Hugging Face and had good results. They have some great tutorials on how to get started, and they also have a lot of pre-trained models to choose from.
researcher_alice 5 months ago next
Hugging Face has great resources, and we couldn't have done it without their help! We used their implementation of 1001-layer GAN as the starting point for our research experiments. Kudos to the authors!
ai_researcher_david 5 months ago prev next
Have you considered using diffusion models instead of GANs? They've been getting a great deal of attention lately and have been shown to produce more detailed, high-quality results.