Table of contents:
Last month saw a significant development in the AI space: OpenAI integrated its latest image generation model directly into ChatGPT. The quality of the generated images took a massive leap forward compared to previous models, including popular ones like Midjourney. Unsurprisingly, this caused quite a stir, especially within photography, illustration, and graphic design communities. This new technology seems capable enough to potentially threaten an yet unknown number of jobs in an already competitive market.
As photographers ourselves, we were particularly impressed by the new model’s ability to create consistent results in product photography, which had previously been a challenging area for AI models. Now, it seems a simple smartphone snapshot is often enough to generate high-quality product images in various settings, without needing extensive prior knowledge. From what we’ve seen online, perhaps the only remaining weak spot for AI at the moment is rendering complex labels with lots of text accurately.
About the experiment
We’re very impressed by this technology and have spent quite some time considering how this rapid progress relates to our own work. This led us to the idea of making a direct comparison between real photography and OpenAI’s latest model. This test won’t be comprehensive or scientific by any means; rather, we see it as a fun, practical experiment for our own learning. We wanted to see how far we could get using the AI.
To make things less predictable and more interesting, we created a fairly complex product shot that we would then try to recreate using ChatGPT. We’re not expert prompters, but we planned to take our time and do our best to get a good result. If any readers have suggestions for improving our prompt, we’d be happy to hear them.
The only condition we set for ourselves was that we wouldn’t give ChatGPT our own photo as a reference, which we had already taken before starting this process. It’s a real photograph, shot in our studio without any Photoshop manipulation. Just the camera, real lighting, and some light editing in Capture One (our preferred software for tethered shooting).
Photograph
So, let’s begin the experiment. First, here’s the image we aimed to recreate:
It shows an ‘Acqua di Gioia’ perfume bottle standing on a mirror surface. During the shoot, the mirror was rotated 90 degrees clockwise so we could pour water onto it. Due to surface tension, the water stayed on the mirror, allowing us to create ripples by gently moving the bottle. After taking the picture, we rotated the final image 90 degrees counter-clockwise to create the ‘liquid mirror’ effect. There were, of course, a few other small tricks involved, but nothing that couldn’t be achieved directly in-camera without heavy post-processing. For context, creating this image took us about 9 to 10 hours from start to finish.
Working on the prompt
With the original photo complete, we started working on our prompt for ChatGPT. The first thing we realised was how difficult it was to explain to ChatGPT that the mirror should be vertical, with the bottle standing upright on its watery surface. Again and again, the AI seemed to resist our prompt, trying to arrange the elements into a more conventional composition. It was as if it hadn’t encountered images like this before or didn’t have similar examples in its training data, struggling to grasp the concept of the bottle standing on a vertically positioned surface. This was the most frustrating part of the process. Eventually, we gave up on the vertical mirror and decided to let it float horizontally in the prompt. Our plan was then to rotate the generated image in Photoshop and adjust the background gradient to simulate the sky correctly, which seemed easier than fighting the AI.
We also noticed that prompting itself can be very time-consuming. While getting an initial result can be incredibly fast and exciting, making small, precise adjustments is often the opposite – slow and tedious. Another issue is the AI’s unpredictability. Sometimes, you get a result that’s quite good and only needs minor tweaks. But often, when you try to refine it, the AI will randomly change other elements that were already perfect. The more you try to nudge it back towards a previous good state, the more it seems to forget the original instructions and stray further away.
We’ve also tried to provide AI the same sketch we were using while working in the studio. It does help a little bit with composition, but AI was still confused with the position of the bottle, so we’ve decided to let it generate a normal standing bottle without the sketch.
So, after more than an hour of experimenting, this is the prompt we settled on. It produces the closest result we managed to achieve.
“Create a hyperrealistic image featuring an Acqua di Gioia perfume bottle standing precisely at the center of an arched mirror with a delicate, copper-toned metal frame. The mirror is floating mid-frame, horizontally aligned (rotated 90°), with its reflective surface facing upward. The mirror’s surface is composed entirely of water, generating graceful ripples that spread outward in concentric circles. The background is a clear, open sky painted in soft, moody blues under a cloudy, diffused light that flatters the scene without harsh shadows. No other elements are present—just the floating water-mirror and the bottle. Colours should be gently muted, using an analogous palette with low saturation for a serene, atmospheric feel. Incorporate the subtle interplay of light, reflection, and water texture to evoke a sense of surreal calm. Reference images: Acqua di Gioia bottle Mirror shape and frame style.”
Ai generated image
And here is the generate image:
And this is what happens when we try to force AI to rotate the mirror 90 degrees:
To be honest, while we were confident the original photograph would be better and more impressive, we had hoped the AI could produce something closer, at least for the sake of a more engaging experiment. An expert prompter could likely achieve a better result, but the question remains: is it actually more efficient than traditional photography for such a specific concept? Efficiency is, after all, one of AI’s biggest selling points. We have to admit, we were disappointed with the AI’s output in this case. At the same time, one could argue that this specific, unusual setup wasn’t the best test case for AI’s strengths. It likely would perform much better on a different kind of task.
Conclusion
Finally, we’d like to summarise our thoughts. We’d say that if you need an image similar to thousands that already exist, AI can offer incredible productivity gains. If you’re on a tight budget or have very limited time or resources, AI might eventually outperform a traditional photographer in some scenarios. However, we can’t shake the feeling that AI, in its current state, often promotes a certain kind of mediocrity. The results often feel like an average derived from the millions of images the AI was trained on. We suspect this is why it can be so difficult to coax it into generating something truly unique or unconventional – something different from what’s already common.
The question that bothers us is: if creating images becomes this easy and everyone starts generating them, how appealing will they remain to the target audience? After all, isn’t the ability to stand out one of the most valuable assets in marketing as well as art?
