Seeing is believing: examples of how new imaging LLMs change the game for better and worse

Use cases for AI have felt elusive at times. Not so for the new autoregressive image making from OpenAI and Google. Getting specific imagery, including correctly written text, was once a total crapshoot. Not anymore.

I’ve been reaching out to clients and colleagues with use cases. See the carousel below.

Some initial observations:

• Just like text, the LLMs are amazing but also make mistakes. The difference is that we can now iterate on specific issues, like spelling (see the digital ad example).
• Prompting strategies are important, just like text-LLM use, in getting what you want the first time. Also like text-LLMs, the technology will undoubtedly improve and require less prompting skill.
• Fraud will be an issue (see car example). Like with text, technologies will have to be developed to detect the LLM images or ensure photos are untouched. I can imagine insurance companies building in specific picture taking to ensure images are real.
• OpenAI has admitted their ‘servers are melting.’ A hack I’ve found is to check both the (iOS) app and the web site for the image rendering. Speed between them varies.