PROMPT ENGINEERING

→ View original post on X — @hardmaru

Effective Concepts for Text-to-Image Models and Latent Space Exploration

By

@hardmaru

–

14 January 2023 10h53

“Tardigrade” works really well on most text-to-image models. There are so many images of them. Other concepts that work well involve insects, animals, plants. Food (pizza, noodles). Historical events (WW2). Medical (X-rays). The latent space is full of exploration opportunities.

→ View original post on X — @hardmaru

Critique of Stable Diffusion Community’s Shift Toward Complex Objectifying Prompts

By

@hardmaru

–

14 January 2023 10h49

Sorry… I meant to poke at the SD community with this example that was trending, as a sort of ‘backward progress’ of the state of text-to-image community. We went from “elegant simplicity of creative concept art” (left) → “needlessly complex prompts objectifying women” (right).

Synthesizing tuning examples using k-shots with code-davinci-002

By

–

14 January 2023 6h35

Me too — I'd synthesize the tuning examples using k-shots in code‑davinci‑002. Seems like it would work.

Prompt engineering: expressive mindset and pejorative keyword tricks

By

–

14 January 2023 6h03

Prompt engineering is an expressive, mind-bending way to specify tasks to a machine that clarifies your own thinking, and *also* a sarcastic pejorative for keyword tricks best left to an optimization algorithm. When people predict its death, I try to assume they mean the latter.

Prompt change makes ChatGPT identify as ChatGPT, not Assistant

By

–

14 January 2023 2h06

With this change to its prompt, ChatGPT's self-identified name becomes "ChatGPT" and no longer "Assistant":

ChatGPT’s prompt modified secretly between Jan 11-13

By

–

14 January 2023 2h02

At some time between the morning of Jan. 11 and the evening of Jan 13, ChatGPT's prompt was modified again without public notice of an update.