News for hackers. Stuff that matters.

A comment on Prompt Engineering Is Hard.

Re:Prompt Engineering Is Hard by d--b on Tuesday October 04, 2022 @08:02AM (#33079157)
It's only a matter of months before someone comes up with a half-text / half-image prompting method for these models.

As in: you'll be able to draw a bunch of circles and arrows with labels in them like "astronaut" / "horse" => "sunset" and some environmental info like "in the style of van gogh" and bang you get a proper image.

[ Parent | on HN ]
Re:Prompt Engineering Is Hard by lifthrasiir on Tuesday October 04, 2022 @08:08AM (#33079201)
In some sense img2img can already do this, I have seen enough artists trying to leverage img2img in their creative process (as opposed to txt2img, which really seems only useful for ideation).
[ Parent | on HN ]
Re:Prompt Engineering Is Hard by willy_k on Tuesday October 04, 2022 @10:15AM (#33080707)
img2img and in/outpainting are already two forms that this idea exists in. Maybe not with textual labels in the image itself, but the concept of having a starting image that a model can modify for a desired change is definitely already a thing.
[ Parent | on HN ]