I spent the last couple of days just doing some prompt engineering with stable diffusion which is mostly trial and error. After that, I had the intention spent time doing EDA on a data set of prompts with user ratings. Turned out,the dataset proved to be not very aesthetic. Just like there is a meta for game there's a meta for prompt engineering and since it was scraped it seemed to be shortly before or after SD became viral, it seemed not to respect best practices in terms of prompts. It feels like playing an early version of Quake at a local LAN almost. I'll probably have to scrape Discord myself for more recent prompts. That or use the Lexica.art API, which i guess ranks them by user preference? I'd have to check but I highly suspect in this case, the wisdom of the crowd is the best way to get good looking prompt. I also started doing EDA on part of the LAION data set.That's 25 GB for the metadata alone, it takes about one hour and a half just to find unigrams, there's also stop so this will take a while. In the meantime, what I'm planning to do today is to start really digging into the theory, make my own notes , probably log back on tonight and write another blog post. I hope to be a little bit more consistent And yeah, that's about it. This post was written with the help of Whisper. This is a great tool for all bloggers. Thanks.