Are foundation LLMs catching up on image gen

I’ll be honest. I am disappointed in the free headshot I got from ChatGPT.

On the other hand, it’s good for business.

Six months ago my team launched PhotoDripAI and we have enjoyed seeing headshots from our platform spread across the LinkedIn profile universe. Yesterday, our good friend Bettina posted a headshot we did for her before launching to the public and it got a lot of love. She’s an incredible human who has given us a lot of encouragement and shared her expertise. If you don’t already follow her you really should. I digress.

My team and I have watched as foundation models eliminate one start-up after another. Many of the casualties were little more than ChatGPT wrappers making API calls so the writing was always on the wall. We hoped our story would be different. There are technical reasons for this I can explain in depth in another post.

When I saw someone post recently about the ‘incredible free headshots’ they got from ChatGPT I wondered if our number was finally up. I was also skeptical because I had tried it months ago with very poor results. I also knew the technical reasons why it was highly unlikely. But, given the way things have gone in the last two years I know better than to assume anything at all about what is possible from one day to the next.

So I decided to set up an experiment. I took a simple prompt and gave ChatGPT an image of myself to see what it would produce. I watched the picture slowly come into focus. And I let out sigh. Partly of relief. Partly of disappointment. Relief because the likeness really was quite terrible. Disappointment because the likeness really was quite terrible. Then I gave the exact same prompt to my platform where a model has been trained on my likeness. By all measures, the images were not only better, they were also closer, way closer, to my likeness.

The second thing I noticed is that the ChatGPT generated images looked very much the same despite me submitting my prompts on separate occasions. The ability of the model to generalize is severely limited when compared with my personal image model. Suppose I wanted to change the background or the style of my clothing. Without being trained on my likeness the foundation llm will struggle to do this.

A final comment on the likeness factor. Even models trained on 30 images of a specific individual will struggle to make a perfect replica of the subject. For me, the likeness is close enough. For some people it is not.

Which camp are you in? Have you tried a headshot generator or ChatGPT? What kind of results did you get? Let me know!

#AI #founder