The GenAI Wine Glass Challenge

I saw this challenge the other day on Reddit. Using the GenAI tool of your choice, create an image of a glass of wine that is filled to rim. I thought it would be relatively simple to do. After all, it's fairly easy to describe what a full glass of wine looks like right? Right?

Unfortunately, I kept getting result that were similar to everyone else. No matter how many times I reworded my prompt, I still kept getting variations of a half full glass. Sometimes they were more full, but they were never filled to the absolute brim.

My prompt for the image above: realistic image of a 12 ounce wine glass filled to the absolute rim with 12 ounce of white wine

Why does this happen? The results are based on what the model was trained on. Think of all the pictures of wine glasses that you have seen. The typical wine glass is 12 ounces, and the typical pour of wine for a serving is around 5 ounces. This results in a wine glass that is half full. The models have been trained with a bias towards the ideal 5 ounce pour of wine.

In short, if you know your results are wrong, you probably need to retrain your model with more or better data. You might have a built-in bias in your data or you might be experiencing an over/under data fit.

Can you get an image with a full glass? If so, what was your prompt?