Loggia dei Lanzi: AI Thermography Enhancement Comparisons through 3D Photogrammetry

Imagine you are a detective trying to solve a mystery hidden inside an old, thick stone wall. You can't just knock it down to see what's inside; that would destroy the history. Instead, you use a special "heat camera" that sees how warm or cold different parts of the wall are. Think of it like a night-vision goggles for temperature: if there's a hidden door or a different type of brick behind the plaster, it might hold heat differently than the rest of the wall, showing up as a ghostly shape on your screen. This is called thermography.

But here's the tricky part: these heat cameras usually take pictures that look a bit fuzzy, like a low-resolution video game from the 90s. To make the pictures clearer, scientists have two main tricks. The first is a hardware trick called microscanning, where the camera shakes just a tiny bit (like a hand tremor) to take many quick snapshots and stitch them together into a sharper image. The second, newer trick is Artificial Intelligence (AI). This is like a super-smart computer program that looks at a blurry picture and guesses what the missing details should look like, "hallucinating" new pixels to make the image look crisp and high-definition.

The big question for historians and architects is: Does making the picture look sharper actually help us find the hidden secrets? Or does the AI just make up fake details that look cool but aren't real? If the computer invents a fake brick pattern, it might confuse the 3D map we are trying to build, leading us to the wrong conclusions about the building's history. This paper dives into that exact question, testing whether these high-tech upgrades are a magic wand or just a fancy filter.

---

The Story of the Loggia and the Heat Camera

The story takes place at the Loggia dei Lanzi, a famous open-air gallery in Florence, Italy, built centuries ago. It's a massive structure with huge stone arches and a vaulted ceiling, visited by millions of people every year. Because it's so old, it has been modified many times. There are likely hidden doors, blocked-up windows, and different types of stone buried under layers of plaster. The researchers wanted to use heat cameras to find these secrets without touching the wall.

In December 2025, a team of scientists set up their equipment in the Loggia. They used a high-end heat camera (the FLIR T1020 HD) to take 161 pictures of the back wall. The weather was perfect for this: it was cold outside (around 11°C) but the inside of the building was warmer. This temperature difference made the hidden features pop out on the heat camera, like steam rising from a hot cup of cocoa on a winter day.

The Great Resolution Race

The team didn't just take the pictures and stop. They wanted to see if they could make the images better using three different methods, and then see which one helped them build the best 3D map of the wall. They treated the images like a race with six different runners:

1. The Native Runner: The original, unedited photo straight from the camera (1024 × 768 pixels). This is the "control" group—what we started with.
2. The Hardware Runner (UltraMax): The camera's built-in trick that uses tiny shakes to combine 16 frames into one sharper image (2048 × 1536 pixels).
3. The Math Runner (Bicubic): A simple, old-school computer trick that just stretches the image to make it bigger, without adding any real new details. This is the "fake it till you make it" baseline.
4. The AI Runner #1 (SwinIR): A smart AI model usually trained on normal photos, not heat ones, just to see if it could guess the details.
5. The AI Runner #2 (DifIISR): A very advanced AI that uses a "diffusion" process (like slowly revealing a picture from noise) designed specifically for heat images.
6. The AI Runner #3 (TongJi/DRCT): The winner of a recent global competition for making heat images sharper, which zooms the image in by a massive factor (8 times bigger!).

The Big Test: Does "Sharper" Mean "Better"?

Here is where the plot twists. Usually, when we see a blurry photo get sharper, we assume it's better. But the researchers weren't just looking at the pictures; they were using them to build a 3D model of the wall. They fed all six versions of the images into a computer program that tries to figure out the shape of the wall by matching points between the photos.

Think of it like a puzzle. If the puzzle pieces have clear, real edges, you can build a perfect castle. If the puzzle pieces have fake edges drawn on them by a computer, the castle might look cool from the front, but it will fall apart or look weird from the side.

The Results:

• The Native Image (The Original): Surprisingly, the original, unedited photo produced the most accurate and reliable 3D map. It had the fewest errors and the most consistent points.
• The Hardware Runner (UltraMax): This one did okay. It was slightly better than the simple math stretching, but not a huge leap. It was the only "enhanced" version that didn't mess up the 3D map too much. It was the closest to the original quality.
• The AI Runners (The "Magic" Ones): This is the big surprise. The AI models that made the images look the sharpest and most detailed actually made the 3D maps worse.
  • The AI models created a lot of "noise." They invented fake textures and details that looked great in a single picture but didn't match up when viewed from different angles.
  • When the computer tried to build the 3D wall using these AI images, it got confused. The "fake" details didn't line up, causing the 3D model to become shaky and inaccurate.
  • The most aggressive AI (the one that made the image 8 times bigger) filled the 3D map with millions of points, but they were all clustered in weird spots, leaving the smooth parts of the wall completely empty. It was like having a million puzzle pieces that only fit in the corners of the puzzle, leaving the middle blank.

What This Means for History

The paper concludes that for the specific job of mapping historic buildings to find hidden secrets, more pixels do not equal more truth.

The AI models are great at making pictures look pretty for a museum poster or a video game. They can guess what a texture might look like. But when you need to measure the wall accurately to find a hidden door, those guesses are dangerous. They introduce "hallucinations"—details that aren't there.

The researchers suggest that if you want to find hidden features in a historic building, it's better to take a good, stable photo with the camera and maybe use the camera's built-in hardware trick (UltraMax) if you have to. But you should be very careful about using AI to "fix" the image, because the AI might be lying to you about what's actually on the wall.

In short: Don't let the computer guess the history. The original, unedited heat map, even if it looks a little fuzzy, tells the truest story of the Loggia dei Lanzi. The AI might make the picture look like a high-definition movie, but for a scientist trying to solve a mystery, the original low-res photo is the only one that doesn't have a fake plot twist.