Matt Wolfe runs the same image prompt 40 times on GPT-5 and GPT-4.0 to see which replicates his headshot better. GPT-5 improves through turn 11 then plateaus, and its final image is closer to the original, though it adds an unexplained grainy black and white filter.
This is my most watched reel to date. So we're going to test the same thing now with GPT5 and see if it does any better. So, same prompt as last time. Create the exact replica of this image and don't change a thing. But I also added this to the prompt. I give you permission to replicate it, and then I use my own headshot, because for whatever reason, the random meme I used last time apparently violated some sort of OpenAI content policy. This time, however, I only ran the prompt 40 times. But you'll see why. All the way through about turn 11, it actually seemed to kind of work, but then it totally plateaued. Turns 11 through 40 are pretty much identical, and I'm not sure what happened between 1 and 11. That's a whole lot different. And here's the final image. Compared to 4.0, this image is way more similar to the starting image. I just don't know what happened with this weird grainy black and white filter that seems to be over the Image. So when OpenAI announced GPT5, they didn't specify if there was any updates to the image generation capabilities. So presumably GPT5 is still powered by the same model that powered 4.0. But some users say GPT5 image generation is notably better. There is a possibility, however, that GPT5 is understanding the prompts themselves way better, and that's why it's creating better images. Now, this isn't a definitive test by any means, but it may show that GPT5 image generation really is better than 4. 0. If you want more quick demos like this, make sure to give me a follow so you never miss a new one.
AI News: Dots, GPT-6.1 Sol, Sonnet 5.5, Gemini 4, and everything you need to know
