Prompt Engineering Is Hard

Re:Prompt Engineering Is Hard by danielvf on Tuesday October 04, 2022 @09:27AM (#33080025)

One fun tip for anyone who wants to understand more how their prompts work - lock down to a single seed, no variations, then change one word at a time in your prompt. Then you can see what effects each word has.

Or you can answer your question about the effects of commas. You can try moving a word or phrase forwards or backwards and see what effect it has.

By working with a single snapshot of randomness, you can see the word effects more clearly.

[ Parent | on HN ]

Re:Prompt Engineering Is Hard by O__________O on Tuesday October 04, 2022 @09:59AM (#33080480)

Agree; to the expand on your points, even trying ultra simple prompt like [black circle on a white background] and then running the output through CLIP to get its text outputs. Then try to narrow the iterations between the two to be predictable and controlled by you.

Obviously understanding the code and data the model is based on is likely best path to better understanding what is going on; for example, last time I checked, Stable Diffusion was limited to 77 tokens.

Ultimately though, possible paths to understanding are like unlimited and just as creative as any other aspect of the topic.

[ Parent | on HN ]