Favorite resources on prompt engineering?
I’m thinking about making a Git repo where we can track all the best incantations.
Maybe I’ll just list them here while I’m thinking.
“Let’s think step by step” hehehe
Trending on Artstation
Using hyphens in this way is something I saw on Twitter first, but now I can’t find the original thread. They were using a white-moth I think?
i don't know if you know this yet but using hyphens often helps it with composition. like "a-discount-store filled with portals-to-other-planets"
— 🎭 (@deepfates) April 25, 2022
Using specific camera lenses and f stops
f3.5 35mm photo (or similar)
— Tim is Making Things In MTL Until Further Notice (@MasterTimBlais) June 22, 2022
Oh, to be clear, I mean for AI generated art and language. Like for DALL-E and GPT-3.
Asking good questions is surprisingly hard lol
Oh yes, this one’s very popular. I find it’s a little too fish-eye or portrait focus or something for me though.
"made in Unreal Engine"
— Max ⛅ (@maxisawesome538) June 22, 2022
Hahaha, I did not know this one. That’s honestly amazing.
Humans can’t do this continuous extrapolation, even though logically it’s what the words imply.
cheap one but seems that chaining “very”s help ie “very very very pretty pic of X”
— cobo (@crinzo_) June 22, 2022
Good overview of the structure of a prompt, I like this.
I feel that there’s some spectrum between stuff like this and stuff like “I am an expert Python developer”… maybe that’s prompt hacking instead?
— Jacques (@JacquesThibs) June 22, 2022
Source on that one:
We found that code models get better when you prompt them with "I'm an expert Python programmer". The new Anthropic paper did something similar, prefixing the model's response with "I’ve tested this function myself so I know that it’s correct:"
— Jacob Austin (@jacobaustin132) April 15, 2022
Maybe there’s a reason this is part of classical rhetoric: you gotta tell the audience who you are, so they know why they should listen to you.
Or maybe it learned this from our pattern of doing that.
Yes I have also found this for math. If you append "I am a math tutor" it starts to answer with higher accuracy.
— Barret Zoph (@barret_zoph) April 15, 2022
This is a really good piece! Lots of necessary detail and visual examples.
I especially like the… ablations? I think they’re called. Where it shows the different prompts and styles in comparison.
Tried writing one up last year. There are probably more complete oneshttps://t.co/Xg5OUapkYz
— Matt McAteer (@MatthewMcAteer0) June 22, 2022
Comparison linked to from there, using VQGAN and a bunch of different keywords.
Fantasy and concept art is in there for sure, but I think it’s especially heavily weighted when you’re using illustrative or “digital art” style cues.
midjourney sometimes likes
— bri (@br___ian) June 22, 2022
4k photo
cgsociety
huperreal
also likes very thematically dramatic stuff like space, heaven/hell, creepy creatures
Oh yes, this is awesome. Being able to “pull” and “push” through latent space, at the same time.
Some models give you an option to provide a negative prompt as well, to get the opposite of what you asked for (basically, another loss function, but negated). Just another tool in the toolbox.
— Peter Baylies (@pbaylies) June 22, 2022
This reel is a great depiction of the process.