Hi friends at HN. Since OpenAI cleared commercial use for their service, I started working on a poetry book illustrated using their imaging model.

I thought it could be an interesting application of the service as there is no storyline from one poem to the other, so there wouldn't be any expectation of continuity from one image to the other. Adding to that, poetry can be very abstract so the set of acceptable images is much larger than when requiring something in particular.

For some details on the creative process:

Most of the time, I would run the poem first as a query and see if the interpretation of the model would yield something interesting. Sometimes, I did have a specific image in mind and I'd try to coax the model to produce that. Much harder than going with its interpretation.

I purchased 3 rounds of credits for the book (barely touching the last one). Usually, I could get something interesting in about 2 queries, but a few took me 5 to 10 tries to get an approximation of what I was looking for. In terms of prompt engineering, I didn't rely on anything too fancy. Mostly appending "digital art", the mood I was looking for, "photorealistic" if I wanted something with more details, "impressionist" for the contrary.

Thinking about the whole process, I think the machine ended up having some influence on the final product. There was one occasion where the picture I got was too good not to use even if it didn't correspond to what I was aiming for, so I adjusted the lines to fit the pic. There was also a few times where one image would inspire the next poem.

Below are some of my favorites (and the preview on amazon also has the first 10 available).

temporal pincer

ghosts of my inverted selves

nudging me forward

https://pbs.twimg.com/media/FaOCif1XkAI2KQe?format=jpg&name=...

---

sea of thousand eyes

and silicon neural nets

exploring frontiers

https://pbs.twimg.com/media/FaPIGu6WAAEVgDL?format=jpg&name=...

---

century's highlight