Testing GPT-4o’s New Image Generator from Photo to AI Recreation

OpenAI recently released the new GPT-4o image generation capabilities, marking a significant step forward in how we can move between text, vision, and creativity. Naturally, I wanted to put it to the test.

Instead of running a standard text-to-image prompt, I flipped the workflow. I called it the Reverse Image Process: starting with a real-world photo, asking GPT-4o to describe it in detail, and then using that description — without the image — to generate a new visual from scratch.

Step 1 | The Original Shot

I took a spontaneous photo with my iPhone while dining at a traditional Tuscan trattoria.

Here’s the shot:

Step 2 | GPT-4o Description

Then, I passed the image to GPT-4o with this prompt:
“Describe the image with as many details as possible.”

Here’s the GPT-4o response:

This image captures the interior of a typical Italian trattoria, with a warm and rustic atmosphere. Here’s a detailed description:

Setting and Architecture:

• Ceiling: Made of dark wood with exposed beams, very high, giving a sense of tradition and architectural solidity.
• Walls: Painted in a warm orange tone, contributing to a cozy and homely feel. They’re decorated with framed paintings, likely depicting Italian landscapes or rural scenes.
• Floor: Covered in terrazzo or graniglia, typical of historic trattorias—durable and easy to clean.

Lighting:

• Hanging from the ceiling are flower-shaped lamps, casting a soft but adequate light.
• Ceiling fans with wooden blades match the rustic style of the place.

Furnishings:

• Tables: Wooden, with yellow paper tablecloths—very common in Tuscan or Emilian trattorias. They’re often used for writing directly on or to maintain an informal vibe.
• Chairs: Wooden, simple and sturdy, matching the tables.
• Tables are neatly set with vertically folded napkins, cutlery, and wine glasses.

Details on the Foreground Table:

• A traditional flask of red wine (with a straw-wrapped base) is clearly visible—an unmistakable symbol of Tuscany.
• Next to it, a glass half-filled with the same wine.
• There’s also a brown paper bag, likely containing bread or other food items.
• A smartphone (iPhone) rests on the table.
• Two glasses (one filled with wine, the other empty), a fork, and a knife are carefully placed.

Guests and Staff:

• A few guests are engaged in conversation, creating a relaxed and friendly atmosphere.
• A waitress in black uniform can be seen in the background, serving or tidying up.

Atmosphere:
The scene conveys a sense of calm, tradition, and conviviality. It looks like one of those trattorias loved by both locals and tourists in search of an authentic experience. The presence of traditional elements like the wine flask and yellow paper tablecloths suggests a simple, genuine, likely homemade style of cuisine.

Step 3 | Reimagining the Scene

Once I had the full description, I opened a new GPT-4o chat and passed it along with a new prompt:

“Create an image based on this description.”

There was no visual input — just the language-based interpretation of the photo. I wanted to see how accurately GPT-4o could reconstruct the moment using only its own descriptive output as reference.

Here’s the result:

Conclusions

I’m genuinely impressed by how accurately GPT-4o was able to translate a real-world photo into a newly generated visual using only its own descriptive output — no reference image involved.

The AI didn’t just generate a generic dining scene. It preserved the architecture, the mood, and even subtle elements like the lighting style, table setup, and objects on the table. The fact that it could reconstruct such a scene purely from language demonstrates a deep visual-linguistic understanding that goes well beyond pattern matching.

What stood out most was the coherence between what GPT-4o described and what it later rendered. The wine bottle wrapped in straw, the yellow tablecloths, the rustic wooden beams, the glass of wine — all these distinctive details reappeared in the AI-generated image, validating how well the model internalised its own interpretation.

This experiment highlights how advanced multimodal AI has become — not just in generating content, but in bridging the gap between visual perception and language. It’s a powerful step toward more creative and fluid human-AI workflows.

There’s a lot of potential here — and we’re just scratching the surface.