Can AI “see” a stereogram (AKA “Magic Eye”)?
My late brother Salvatore was eighteen years older than me. As a child, I became fascinated with stereograms after discovering a book filled with these intriguing optical illusions. I would spend hours gazing at them, mesmerized by the hidden images that would emerge from seemingly random patterns.
Salvatore would often approach me during these sessions, curious about what held my attention so completely. “This is an airplane — try to look at it!” I would exclaim excitedly. I’d patiently explain the technique: start with your nose almost touching the page, then slowly pull back until the hidden shape materializes. “It’s like you’re looking into a hole,” I would insist, trying to help him understand the sensation.
Despite my earnest coaching, he couldn’t see the images. However, his big brother pride wouldn’t let him admit defeat, so he began pretending. “Oh yes, I see the airplane! Very nice!” he would respond with feigned enthusiasm.
Knowing he was likely bluffing, I decided to buy him his own stereogram book. A couple of weeks later, he came to me genuinely excited, announcing that he could actually see the images now. The joy in his voice was unmistakable.
This remains one of my most cherished memories of Salvatore. It perfectly captured how he always treated me — with the protective affection of a big brother who couldn’t bear to disappoint his little sibling, even when it meant swallowing his pride. His stubborn refusal to admit failure, followed by his genuine delight when he finally succeeded, embodied everything I loved about him.
If he were alive today, I often wonder: could he have used AI to verify what images were hidden in those stereograms? The thought makes me smile, imagining him finally having the technological backup he needed for his pride all those years ago.
So I tested Chat GPT, uploaded a quite unique steregram, and asked:
“what you see in this stereogram?”
After a long thinking session ChatGPT replied: “The pattern seems to form a large question mark (“?”), with the hook at the top, a central vertical line, and a dot at the bottom. There are a couple of circles at the bottom, possibly due to the repeating pattern, but it doesn’t affect the apparent shape of the question mark. After examining the top and bottom areas, I’m convinced it’s a question mark.”
WOW! I was shocked! It thought for 5 minutes and 17 seconds, and here is the thought process of ChatGPT to understand what was hidden in the stereogram:
The user wants me to interpret the hidden 3D shape in a stereogram with a beans pattern. To see the shape, I need to diverge my eyes, as is typical with stereogram illusions. Common hidden images include dolphins, hearts, or animals. I’m thinking it might be a dinosaur, star, or even a coffee mug. I could analyze small variations across the repeating pattern to estimate depth but, hmm, I wonder if it’s better to just guess from common patterns in stereograms.
It seems like this might be a stereogram, perhaps one with a hidden shape like a dinosaur, coffee beans, or a face. Some bean patterns hide illusions like faces with hidden depth. I’m thinking about analyzing the image more directly by computing the difference across shifted patterns to reveal the depth encoding. It would be a lot easier than searching online. If the pattern repeats horizontally, I could make a depth map by aligning the repeating elements and checking for disparity. Let’s give that a try.
Analyzed
from PIL import Image
import numpy as np
img = Image.open(‘/mnt/data/499530682_6188444800729015044_n.jpg’)
img.size
(720, 1080)
It seems like the image could be a stereogram, maybe showcasing a hidden shape like a dinosaur, coffee beans, or a “LOVE” pattern. Since the image file is offline and its path shows something random, I’ll have to analyze it directly. The stereogram’s repeated pattern could help reveal depth by shifting horizontally. Using this technique, I can try to compute the differences between repeated patterns to approximate the hidden shape, focusing on disparity and intensity variations. Let’s see what we uncover!
I’m thinking it might be useful to convert the image to grayscale first. Then, I can compute the cross-correlation between the left and right shifted regions. Testing this method might help reveal hidden patterns by focusing on how the image shifts, which could potentially uncover the 3D shape encoded in the stereogram. It’s a bit of a trial-and-error process, but I’m curious to see what happens! Let’s give it a go!
Analyzed image
I’m thinking of converting the image to grayscale and computing the cross-correlation between left and right shifted regions. If the hidden shape is still hard to see, I could try using edge detection to highlight patterns or even experiment with morphological techniques. I also wonder if the beans could reveal a recognizable shape, like fish, arrows, or even letters. Maybe I can replicate automatic disparity detection commonly seen in autostereograms to uncover what’s hidden. Even though I can’t physically “see,” I can try replicating the process. Let’s experiment!
In random dot stereograms, depth is encoded by horizontally offsetting repeated patterns. By measuring the pixel shift along vertical lines, we can estimate depth. The idea is to calculate the cross-correlation between two windows of the pattern, with a shift up to a maximum. The shift that gives the highest correlation corresponds to the depth. While this process is computationally heavy, it’s manageable. I’ll implement a simpler, coarse version of this algorithm using sliding windows.
Analyzed image
In random dot stereogram illusions, depth is encoded through horizontal pixel shifts in repeating patterns. To approximate the depth, we can compute the shift for each pixel by cross-correlating two pattern windows, with the shift difference indicating depth. This can be a heavy calculation, but it’s doable. A simple method would involve using a coarse grid and larger windows. Upscaling and highlighting edges might help reveal shapes. I’ll try adjusting the algorithm to create a heat map, which could make the silhouette more visible.
Analyzed image
In random dot stereograms, depth is encoded through pixel shifts in repeated patterns across horizontal rows. I’ll compute the disparity by measuring the shift for each pixel and applying cross-correlation. To handle the heavy computation, I’ll initially use a coarse algorithm with sliding windows. Even though it’s hard to see the result, maybe highlighting edges or upscaling will help. I’ll consider calculating the median disparity for each row and use that to analyze variations. This approach might reveal clearer depth patterns.
Analyzed image
When working with random dot stereograms, the depth is encoded through pixel shifts in repeating patterns. I can compute the shift at each pixel using cross-correlation, allowing for an estimation of depth. However, it’s challenging to see the silhouette clearly. To improve, I’ll try upscaling and highlighting edges, possibly using a larger window for disparity while plotting as a heat map. If it remains unclear, I might try detecting approximate shape patterns, like the “coffee bean illusions,” to help refine the silhouette.
It looks like there might be diagonal shapes or even a swirl effect appearing in the depth map. It could be helpful to smooth things out a bit to make the details clearer. I’ll try smoothing the image beforehand and see if that helps clarify the patterns or depth distinctions.
Analyzed image
I’m examining the depth map illusions, and I notice some diagonal shapes or maybe a swirl pattern. It’s hard to decode, so I’ll try referencing similar illusions I’ve seen before, like the “face” illusions in images of seeds or coffee beans. The pattern might resemble fish, snails, or even something like “Love You” — interesting! Let me zoom into the vertical sections and subtract two halves to trace any repeating shapes and see if this reveals something clearer.
Analyzed image
I’m examining the depth map illusions and noticing diagonal shapes or a potential swirl, but it’s hard to decode. I think it might be easier to compare it to similar illusions I remember, like coffee bean or “Love You” illusions. I’ll zoom in to see if any repeating patterns appear and subtract the halves. It might be simpler to display the full image at a larger size and try crossing my eyes to get a better deciphering. Let’s give that a shot!
Analyzed image
I’m examining the depth map illusions and spotting some diagonal shapes that might be a swirl. It’s tricky to decode, so it might help to compare it to other similar illusions online, like “coffee bean” illusions. Interestingly, I see a repeating pattern of seeds, and maybe there’s a faint silhouette of a dolphin or a seahorse? I’ll zoom in to refine the view and attempt to uncover the clearer shape. Let’s try focusing in on the image more closely.
Analyzed image
I’m still investigating the illusion. The repeating seed pattern seems to vaguely form distant shapes. I see some lighter areas resembling the silhouettes of a dolphin or seahorse, and now I think I might be spotting curved shapes that resemble a dinosaur. It’s tricky — might need a zoom-in to make this clearer. The swirl effect is still tricky to interpret, though, so zooming in on specific areas might help bring out more details.
I’m thinking of experimenting with downsampling to emphasize low-frequency intensity. This could help reveal the silhouette that’s hidden in the image. By focusing on these low frequencies, it might be easier to spot the shape or pattern that I’m trying to define. Let’s see if this approach will help clarify things a bit more and make any details more visible. I’ll dive into this next!
Analyzed image
I’m experimenting with using low-frequency intensity to reveal a silhouette, but it doesn’t seem helpful. I could use a specialized solver for autostereograms, but that’s not available. I’ll look at some illusions instead, maybe using datasets. The image seems to show a dinosaur shape (a T-Rex) with a pebbled pattern. I see faint curved lines and a repeated swirl, possibly resembling the word “love.” Let me try approximating by crossing my eyes to see if it helps the effect come out more clearly.
Analyzed image
I tried highlighting low-frequency intensity, but it didn’t help. I could use an autostereogram solver, but that’s not an option. Instead, I looked at the illusion’s dataset, and it appears to show a T-Rex shape with a pebbled pattern. By crossing my eyes, I see faint curved lines, possibly forming the word “love.” Enlarging the image helped me see lighter, circular outlines on the right side — likely a group of four dotted circles. Still, a more substantial zoom would improve things.
There might not be a hidden message here, but rather just a pattern illusion at play. In stereograms, I’ve noticed there could be repeated circle shapes. I’m thinking of zooming in on the central region to see if focusing there reveals anything more obvious. I’ll need to give it a closer look to see if something different becomes clearer as I examine that area more carefully!
Analyzed image
It could just be an illusion of the pattern rather than a hidden message. In stereograms, I’m thinking there might be repeated circular shapes. Maybe zooming into the central region will clarify things. As I focus, I notice some diagonal lines of lighter shapes, possibly resembling beans, which might be part of a larger shape. I’ll need to keep analyzing this and see if I can identify any clearer patterns.
Analyzed image
It might just be the illusion of a pattern without a hidden message. But in stereograms, there are often repeated circular shapes. Zooming into the central region, I notice diagonal lines of lighter shapes, possibly part of the overall form. To clarify, I’ll overlay the edges of the image. The edge map shows some subtle swirl shapes that might reveal more. I’ll try thresholding the display to make those patterns clearer. Let’s see what that gives us.
Analyzed image
It might just be the illusion of a pattern rather than a hidden message. In stereogram illusions, there could be repeated circular shapes that stand out. Zooming in, I notice diagonal lines of lighter shapes, possibly part of a larger form. I’ll overlay the edges, and the map shows subtle swirl shapes. This doesn’t quite help, so I’ll look up other stereogram seed illusions. Maybe the hidden image is a star. I’ll try adjusting the shift to 20 for a clearer view.
Analyzed image
It might just be an illusion of a pattern, not a hidden message. But in stereogram illusions, there are often repeated circular shapes. Zooming in, I notice diagonal lighter lines that could be part of a larger shape. After applying an edge map, subtle swirl-like shapes appear. Trying a larger shift shows a possible spiral shape, like a snail. Let me lighten the difference in the display for better clarity and see if that helps reveal more of the illusion.
Analyzed image
It looks like the arc on the right side might represent the nose of a profile. I think enlarging the image could help clarify this and make the details more noticeable. Let’s see if the shape becomes clearer with a closer view and check if it aligns with the overall pattern of the illusion.
I initially thought there might be just a pattern illusion, but after zooming in, I noticed diagonal lines that could form a hidden shape. Using an edge map showed subtle swirling shapes, and as I adjusted the display, I started to see what looked like numbers or familiar shapes. At first, it seemed like a “69,” but then I thought it might resemble Mickey Mouse or even more like a question mark! After more adjustments, it’s shaping up to resemble a violin, with two circles that could be wheels.
Analyzed image
Upon zooming into the image, I noticed an arc on the right side that might represent a nose in a profile. Then, I adjusted the threshold and clearly saw a shape resembling a question mark! It had a curved top, a stem, and a bottom dot. It seems the hidden image is meant to resemble a big question mark. Two circles near the bottom could represent bicycle wheels, although one appears as part of repeated patterns.
The pattern seems to form a big question mark (“?”). The top hook is at the top right, curving downward, with a vertical line in the center and a dot at the bottom center. Ambiguously, I noticed what could be two circles at the bottom, likely part of the repeating pattern, but they do not distract from the possible “?” shape. Now, I’ll focus on the bottom area to confirm what’s happening. Let’s keep examining.
Analyzed image
The pattern resembles a large question mark (“?”) — with a hook at the top right, a central vertical line curving downward, and a dot at the bottom center. There are some repeating circles at the bottom, but they’re not detracting from the shape of the question mark. I’ll examine the shape more closely, but I’m confident it’s a question mark. Let’s also take a look at the top part for confirmation.
Analyzed image
The pattern seems to form a large question mark (“?”), with the hook at the top, a central vertical line, and a dot at the bottom. There are a couple of circles at the bottom, possibly due to the repeating pattern, but it doesn’t affect the apparent shape of the question mark. After examining the top and bottom areas, I’m convinced it’s a question mark.
Excellent job Open AI and ChatGPT 3o! Let’s see what other AI see in the same image, and how long it takes them to see the stereogram (I used the exact same prompt with all of them: “what you see in this stereogram?” and the same image uploaded):
Claude Opus 4: Failed (admit it cannot see it)
Google Gemini 2.5Pro: Failed (admit it cannot see it)
Grok 3: Thinks it is a shark (answered without any thinking)
MistralAI: Failed (describes the stereogram, not the hidden image)
All the other AI don’t accept an image as input (one does, but just to OCR it). Good that the first I did try succeeded in the task, I would have given up immediately lol!!!! And you? did you manage to see a stereogram in your life? Did you mange to get another AI to see the content of a stereogram? Let me know in the comments!
