The day I realized my prompt engineering was completely backwards

I spent 6 months writing these long, detailed prompts for DALL-E 3 thinking more words meant better results. Then a guy at a meetup in Austin said keep it short and let the model breathe. I tried it with a 10-word prompt instead of 50 and got exactly what I wanted on the first try. Anyone else find that shorter prompts work way better than they expected?

2 comments

2 Comments

anthonymurray23d ago

You might be onto something about giving the model room to work, but I think the real trick is figuring out when long prompts actually help. For photorealistic images, short prompts work great because the model has a strong baseline. But I've found that with more abstract or surreal concepts, a few extra words are needed to steer the model away from generic results. It's less about prompt length and more about knowing which parts of your idea the model can fill in on its own.

johnson.daniel22d ago

Yeah, @anthonymurray that's pretty much what I've noticed too. The model already knows what a photorealistic cat looks like, so you're just wasting words describing it. But when I tried to get a "clock melting over a desert landscape with some kind of fish floating in the sky" it just kept spitting out generic stuff until I added a few more specific descriptors. The key is knowing where the model's training data is weak and filling in those gaps. Basically, you're just helping the model out where it needs it most.