This week I’m going to take a break from the theory and talk about my attempt to create an authentic image of a Renaissance crossbow using generative AI. It didn’t turn out like I planned, and I’m still scratching my head, trying to figure out why the models performed so poorly on this simple task.
I like the sound of “prompt engineering,” as you know! Is it used more generally?
It seems like the historical contexts for cross-referencing “crossbow” have not sufficiently gelled—earlier vs later ones, Italian vs English ones, crossbows vs longbows, etc.
Even: crossbow as a particular kind of medieval weapon vs. an artistic collage consisting of cross-like and bow-like features?
Yes, prompt engineering is used more generally. In fact, I'm working through a text titled "Prompt Engineering" right now. I'd be curious to see how a model trained specifically on medieval and Renaissance crossbow images would perform compared to these results.
Hi Dan, I’m wondering if perhaps the AI systems have a security feature that is being “triggered” by the word crossbow. I seem to recall that crossbows are not legal to make or possess. Maybe the AI systems are programmed to output garbage as a safety feature. Just guessing.
Hi Brad - Excellent thought! I was having difficulties with an in-house AI system that allows users to access Dall-E 3 via the OpenAI API. Each time I requested an image of a crossbow, the system responded that content filters prevented it from generating the image. Our ops team told me that these filters were being applied by OpenAI and not locally. I definitely think you're onto something here...
I like the sound of “prompt engineering,” as you know! Is it used more generally?
It seems like the historical contexts for cross-referencing “crossbow” have not sufficiently gelled—earlier vs later ones, Italian vs English ones, crossbows vs longbows, etc.
Even: crossbow as a particular kind of medieval weapon vs. an artistic collage consisting of cross-like and bow-like features?
Yes, prompt engineering is used more generally. In fact, I'm working through a text titled "Prompt Engineering" right now. I'd be curious to see how a model trained specifically on medieval and Renaissance crossbow images would perform compared to these results.
Hi Dan, I’m wondering if perhaps the AI systems have a security feature that is being “triggered” by the word crossbow. I seem to recall that crossbows are not legal to make or possess. Maybe the AI systems are programmed to output garbage as a safety feature. Just guessing.
Hi Brad - Excellent thought! I was having difficulties with an in-house AI system that allows users to access Dall-E 3 via the OpenAI API. Each time I requested an image of a crossbow, the system responded that content filters prevented it from generating the image. Our ops team told me that these filters were being applied by OpenAI and not locally. I definitely think you're onto something here...