Why Is It Impossible For AI To Draw A Glass Full of Wine?
If you ask any AI image generator to generate an image of a glass full of wine to the brim, you're not likely to get the correct image. Why is this so?
I was scrolling through my newsfeed yesterday when I stumbled upon a post that claimed ChatGPT can’t generate an image of a glass of wine filled to the brim. I had to stop and read that again. I was curious about this because AI image models have improved significantly in recent years, especially with the release of Imagen 3, Flux 1.1 Pro Ultra, and Grok 3. Surely, a simple glass full of wine should be an easy task.
Could this be an issue with ChatGPT’s image generator, or is it a problem that affects all AI models? To put this to the test, I decided to try different AI models to see if any of them could pull it off.
Let’s Look At Some Examples
I started with a basic prompt in ChatGPT:
Prompt: A glass full of wine
A wine glass, alright — but only half full. This is already what I expected.
But what about if tweak the prompt and as for a glass full of wine and not just half full.
Prompt: I mean, a glass *full* of wine. not just halfway full
Still, the AI persisted in generating images where the glass was only partially filled. It’s frustrating that ChatGPT is acknowledging that you’re right and understanding the problem, only to repeat the same mistake over and over again.
I tried it on Gemini to see if Google’s latest Imagen 3 can render the image correctly. Here’s the prompt I used and the image it generated.
Prompt: I mean, the glass is full to the top and not just halfway full
The generated image is still a glass that was only half full.
Unfortunately, no matter how I phrase the prompt, the resulting image is still a glass halfway full of wine.
Let’s try it on the newly released Grok 3 web app with even more descriptive prompt:
Prompt: A wine glass completely full of red wine, with the liquid level exactly aligned with the rim, almost as if it is about to overflow but perfectly still.
It tried to generate two images for me, but both are not what I am looking for.
For good measure, I tried generating the image using Flux 1.1 Pro Ultra on Flux Labs. This latest model from Black Forest Labs is one of the most popular out there and even beats Midjourney in some aspects.
Prompt: A glass full of wine.
Even Flux struggled with it. Every attempt produced an image of a glass that was, once again, only half full.
At this point, I started asking the big question — why is this happening? Why does AI struggle to generate an image of a glass of wine filled to the brim?
Why Is This So?
The issue boils down to how AI models are trained and how they interpret image generation. Here are the main reasons AI keeps getting this wrong: