Historically, prompting ChatGPT (or DALL-E through ChatGPT) for variations or subtle modifications of a single image results in a sort of image degeneration. Not only does the image not get reproduced with the slight variations, but it gets further and further away from what you originally requested. This is something that has been noted on Janelle Shane’s blog, AI Weirdness (in fact, it’s the last thing she posted there). The entire post is a great, wild read. In it, she notes:
Asked to remove a person from a shot or fix perspective errors, the prompters would return completely different art instead, or with other weird changes that didn’t fit the brief.
This has been my experiences as well. Or maybe…until now?
Last week was the 2024 Presidential Debate between Kamala Harris and Donald Trump. Because I seemed to have found a way to circumvent the restrictions about creating images that clearly depict real people (see my last post), I thought I would see if it was possible to do the same thing with the debate candidates. It failed fantastically, and the closed I could get were these images:
This might be changing after the most recent OpenAI update. A preview release of their newest model is available, so I thought I would see if that could generate the image I was looking for. That model can’t do images at all, but it was able to produce a decent prompt that I fed back into GPT-4o. This is where I started:
Now, this isn’t exactly what I asked for. The prompt was:
In a satirical and humorous portrayal of a fictional 2024 presidential debate held on September 10, two candidates stand behind their podiums on a brightly lit stage. The first candidate is dressed in an eye-catching bright purple suit. His wild blond hair forms a massive combover that flies in all directions, adding to his animated appearance. His eyes are comically wide and bulging as he is captured mid-scream, conveying intense emotion.
Opposite him, the second candidate wears a cool blue suit and looks utterly aghast. Her eyes are cartoonishly popping out, emphasizing her astonishment at the situation. The backdrop features an array of bright lights and a prominent banner that reads “2024 Debate,” setting the scene of a high-stakes event.
In the audience behind them, a comical assortment of shocked faces reacts to the exaggerated and ridiculous behavior of the candidates. Their expressions range from disbelief to amusement, amplifying the satirical tone of the scene.
Obviously the second candidate wasn’t female. But, enjoying the cartoonish, bulging eyes character, I decided to just focus on him and simplify things. This is where things get interesting. It produces a character, similar angle, but no longer with bulging eyes and he was far younger than I wanted him to be (this is supposed to be Trump, or, at least, Trump-adjacent). So, I proceeded to tell the AI to make the character older and with more and more exaggerated bulging eyes. What follows is the results of that exchange in which you can clearly see the same image being subtly tweaked.
This is something I haven’t seen DALL-E/ChatGPT do before, and it goes against what Janelle Shane observed several months ago in her post. Each time, I only gave the minor-est of instructions (“older”, “eyes bulging even more”, “he should be wearing a pin that says ‘winner’”, etc) and it kept an almost identical image throughout, just making those slight changes.
This suggests that AI/ChatGPT is getting better at artistic revisions, which is a very new thing that I found purely by accident.
Leave a Reply