Make It Possible
AI: I tested 50 times. Each generated image looked different.
I tested at least 50 times and AI still gave me images that looked like they were drawn by a sleep deprived medical student.
I said I would share AI limitations and use in my adaptation to hospital and equipment planning. Viola, here we go.
AI is inconsistent. That was the first thing I learned trying to adapt it to my work.
I was generating comic strips for my field stories. A comic so readers don’t need to read a passage of 500 words to grasp one life instance.
What I got was various across different generations. A high school science textbook once. Heavily annotated like an aged science magazine another time. Made up device names once. Made up regulation numbers the other time.
I tested at least 50 times and tweaked the wording in the prompt each round.
Eventually I found things that helped. Adding “newspaper strip” helped it understand the format. Writing “IMPORTANT: do not make up device names or regulation names” reduced the hallucinations. Adding “no annotation” stopped the labels.
But the images were still inconsistent. Each generation looked different. Because it is almost impossible to manually type the exact same checklist every time and also adapt it to a specific story.
The only thing that worked was using a language AI to take the story, tidy it up, and turn it into a prompt that already included the full checklist. Then refeed it into an image AI.
One AI prepares the prompt. The other generates the image.
I want a systematic process so I baked the prompt into my AI agent. But if you are doing a one-off exercise, using a generic LLM works well. Gemini works particularly well as it can generate the image in the same chat interface.
This saved me from numerous retries, long waits just to get images that look like they were drawn by a sleep deprived medical student.
You may try it now. Instead of asking the AI to generate an image directly, chat with it for a minute to set the prompt up first.
This goes the same as setting up text format and table content if you are doing reports or other docs.
Someone you know has been through this? Forward it to them.
