Can your AI image generator create an image of a full glass of wine? Try it and I bet your answer will be no.
This is actually something I tried myself with Leonardo AI and I just couldn't get the tool to give me an image of a glass filled with wine to the tip of the glass.
It's always half full or trying to look more than half or something weird but never a full glass of wine.
I understand the rules of sharing AI images on Hive so I avoid them. However, to proof my point I took screenshots of my own generated images from Leonardo AI.
The prompt I used was - Image of full glass of wine
You would think a prompt this precise and specific would give the results but it couldn't do it.
Open AI noticed it a while ago and have been working on their newest AI image models. They have successfully been able to improve the model to actually generate that image.
I tried using a prompt for Chatgpt and this was my results.
Also screenshot of my generation
Now guys, I took the exact same prompt I used to generate this full glass of wine in Chatgpt and sent that prompt to Leonardo AI.
The Prompt - Create Image of a full glass of wine. I want the wine to be full to the tip of the glass. A full glass
Below is a screenshot of the results I had from Leonardo AI, lol complete and total failure.
AI being able generate a completely full glass of wine might sound a bit trivial to you. You could easily brush it off and say, what's the significance of that or why go through the hassle to get AI to do something like that?
It's not just about image generation, it's a powerful sign of how good these AI systems are now able to process reality.
You might think perhaps it was a bug in the AI system, but no it wasn't a bug. Every model was suffering from this.
It was a clear limitation in how AI perceived physical properties. For years, image generators failed this very simple task because they relied on patterns in their training data. The issue came from how much of partially filled wine glass images these AI models was trained on.
Almost every real life image of a glass of wine was almost empty glass. Just a little in the glass, never full.
You think I'm lying? Fine... Go to Google and just search wine glass or glass of wine. Tell me how hard it was to spot one glass, just one which was actually full, lol.
@meno you wanna try this, hahaha
Since most real world photos of wine glasses show them partially filled, AI models couldn’t seem to break free from that pattern. That was enough proof that the models weren’t thinking at all. They were just repeating what they've been trained on.
And that's why you shouldn't use AI articles for your presentation, it's full of Plagiarism haha.
Thankfully the latest Open AI GPT-4o has kind of fixed this problem and if you observed the image I generated with Chatgpt, it's a full glass of wine.
This is not an improvement in rendering visuals but a great sign that AI is starting to grasp abstract concepts. It can now generate complex images that feels more real world than previous models could.
Better AI generated images actually means better communication skills of the models. This is so important because if we are looking forward to making AI handle complex jobs that only people could, it probably needs to understand commands and give more accurate results otherwise it's no good for the job world.
Just think about it, technical diagrams engineers and medical doctors use, instructional graphics, even something as basic as an accurate restaurant menu photo, all these rely on visual consistency and accuracy.
If AI can finally get a wine glass right, maybe it can get everything else right too. That’s the real breakthrough we're happy about. Not wine, but understanding.
If you've got any AI image generator that can produce a full glass of wine let me know in the comments, I'll look it up and try some prompts.
Oh and people are testing chatgpt out to get the full glass of wine, you should try it too.
Posted Using INLEO