Post 6: The evolution of generative models for images 🕰️

It’s been about 8 years since the first neural networks were used to generate images. The first ones were only a few pixels in resolution and in black and white, but as new architectures were developed (convolutional neural networks, generative adversarial neural networks, transformers, diffusion models), we can now generate high-resolution images that are quite plausible

• Here’s a curious page to see how good we are at detecting whether an image was made by AI or by a person.
• As well as an image bank made by AI

Now there are several websites that test you in detecting images made by AI. Out of 100 images, I was able to correctly detect 67 if they were made by artists or by artificial intelligence 🤕 … That’s almost a coin flip. If you want to try it out, click here. It’s not just an image test, you can also do it with texts.

• http://aiorart.com/

In my opinion, the most promising generative AIs for the world of proteins are:

1. ProtGPT2 (generation of protein sequences)
2. ProteinMPNN (generation of sequences given a structure)

And although undoubtedly a protein is more complex than an image, how long will it take to be able to synthesize proteins with a specific function as easily and well-folded as this? 🧐 Perhaps the starting year will be the publication of AlphaFold2 (2021).