Perceiving animacy in 'identical' images

By utilizing diffusion-based "visual anagrams" to isolate animacy from lower-level visual features across seven pre-registered experiments, this study demonstrates that the human visual system inherently represents animacy itself, which subsequently structures visual working memory and guides attention.

The Gist

Imagine you are looking at a picture of a dog. You instantly know it's alive, it has feelings, and it might chase a ball. Now, imagine you look at a picture of a boot. You know it's just an object; it's dead, it has no goals, and it won't chase anything.

For decades, scientists have known that our brains treat "living things" (animate) and "non-living things" (inanimate) very differently. We spot a tiger in the bushes faster than a rock, and we remember a puppy better than a chair.

But here's the problem:
Scientists have always struggled to prove that our brains are reacting to the concept of "life" itself. Why? Because dogs and boots look totally different. Dogs are furry, have tails, and are curved. Boots are stiff, have laces, and are angular.

So, when a scientist says, "Our brains react faster to dogs," a skeptic can always argue: "No, you're just reacting to the fur and the curves, not the fact that it's alive!" It's like trying to prove that people prefer chocolate ice cream over vanilla, but the only time you offer chocolate, it's served in a fancy gold bowl, and vanilla is served in a cardboard box. You can't tell if they like the flavor or the bowl.

The "Visual Anagram" Magic Trick

To solve this, the researchers (Tal Boger and Chaz Firestone) used a clever trick from the world of AI called "Visual Anagrams."

Think of a Rubik's Cube. If you twist it, the colors on the faces change, but the plastic cube itself is the exact same object.

Now, imagine a single image that is a dog when you hold it upright. But if you turn that exact same image 90 degrees on its side, it suddenly looks like a boot.

• Same pixels: The image is identical.
• Same texture: The "fur" of the dog is the same "leather" of the boot.
• Same shape: The curves are the same.
• The only difference: The meaning (animacy) changes because of the rotation.

It's like a magic trick where a picture of a duck turns into a rabbit just by tilting your head. The researchers used this to create a "perfect control" experiment. They could change the animacy (from alive to dead) without changing any of the visual details like shape or texture.

The Experiments: What Did They Find?

The team ran seven different experiments to see if our brains still reacted to the "life" part, even when the "look" part was identical.

1. The Memory Game (Working Memory)

• The Setup: Participants looked at a screen with five pictures (like a horse, a car, a rabbit, etc.). Then the pictures disappeared. A new picture appeared, and they had to click on what changed.
• The Trick: Sometimes the change was "boring" (a rabbit turned into a dog—both alive). Sometimes it was "wild" (a rabbit turned into a boot—alive to dead).
• The Result: Even though the boot and the dog were the same image just rotated, people spotted the "alive-to-dead" change much faster. Their brains were screaming, "Hey! Something is alive now, and then it's not!" even though the pixels hadn't really changed.

2. The "Where's Waldo?" Game (Visual Search)

• The Setup: Participants had to find a specific target (like a duck) hidden among other objects.
• The Trick: In some rounds, the duck was surrounded by other ducks (all alive). In other rounds, the duck was surrounded by boots (alive surrounded by dead).
• The Result: People found the duck much faster when it was surrounded by boots. Their brains were using the "alive vs. dead" difference as a shortcut to find the target, ignoring the fact that the visual shapes were identical.

3. The "Blob" Check (Ruling out the rotation)

• The Skeptic's Question: "Wait, maybe it's not about life. Maybe it's just that the image is rotated, and our brains hate rotated things?"
• The Fix: They took the images and turned them into simple, indistinct black blobs (silhouettes). These blobs kept the same rotation and shape as the original images, but you couldn't tell if they were a dog or a boot anymore.
• The Result: When the images were just blobs, the "search advantage" disappeared. People were no faster finding the target. This proved that it wasn't the rotation causing the effect; it was the perception of life.

The Big Takeaway

This paper is like finally proving that the "Gold Bowl" wasn't the reason people liked the ice cream.

By using these "visual anagrams," the researchers showed that our brains have a special, built-in detector for "life." We don't just see shapes and textures; we instantly extract the concept of "animacy" (is it alive?) from the visual world, even when the visual clues are identical.

It suggests that recognizing "life" is so important to our survival (to spot predators or friends) that our brains prioritize it above almost everything else, processing it as a fundamental feature of vision, not just a complicated thought.

Technical Summary

1. Problem Statement

The distinction between animate (e.g., animals) and inanimate (e.g., objects) entities is a fundamental component of human cognition, influencing visual attention, memory, and neural organization. However, a persistent methodological challenge has prevented researchers from conclusively determining whether the visual system processes animacy itself or merely its lower-level visual correlates (e.g., shape, curvature, texture, spatial frequency).

In natural stimuli, animate and inanimate objects differ simultaneously in high-level category and low-level features. Previous studies attempting to control for these features have been limited, often leading to the conclusion that effects attributed to animacy might actually be driven by differences in curvature or shape. There has been no prior method capable of varying animacy while holding nearly all other visual features constant.

2. Methodology

The authors introduce a novel experimental paradigm using "visual anagrams" generated via diffusion-based generative AI (specifically leveraging the model from Geng et al., 2024).

• Stimulus Generation: The researchers created static images that appear as one object in one orientation and a completely different object when rotated 90°. For example, a single pixel array appears as a "dog" (animate) in one orientation and a "boot" (inanimate) in another.
• Control Mechanism: Because the two interpretations are the exact same image (just rotated), they share identical pixel values, texture, spatial frequency, and shape properties. The only variable that changes is the perceived category (animacy) and the orientation.
• Experimental Design: The study consists of seven pre-registered experiments involving human participants recruited via Prolific.
  • Experiments 1–2 (Visual Working Memory): Participants viewed an array of 5 anagrams. After a delay, one item changed. Changes were either cross-category (animacy changed, e.g., rabbit $\to$ boot) or within-category (animacy preserved, e.g., rabbit $\to$ dog). Participants had to identify the changed item.
  • Experiments 3–4 (Visual Search - Target Present): Participants searched for a previewed target within an array of 6 anagrams. Arrays were either "mixed animacy" (target differs in animacy from distractors) or "uniform animacy" (target matches distractors).
  • Experiments 5–6 (Visual Search - Present/Absent): A standard present/absent search task was used to replicate the search advantage in a more conventional paradigm.
  • Experiment 7 (Control for Orientation): To rule out the possibility that the results were driven solely by the change in orientation or aspect ratio, the authors created silhouetted versions (convex hulls) of the anagrams. These silhouettes retained the orientation and aspect ratio of the original stimuli but removed all internal texture and shape details, rendering them unidentifiable as animate or inanimate.

3. Key Contributions

• Novel Stimulus Paradigm: The paper establishes the use of diffusion-generated visual anagrams as a rigorous tool for isolating high-level perceptual properties (animacy) from low-level visual features. This offers a level of experimental control previously unattainable in vision science.
• Disentangling High-Level from Low-Level: The study provides the first direct evidence that the visual system extracts animacy as a distinct feature, independent of shape, texture, and curvature.
• Resolution of Controversy: The findings challenge the prevailing view that animacy effects in visual search are merely artifacts of low-level feature differences (such as curvature), demonstrating instead that animacy is a privileged, high-level representation.

4. Results

Across all seven experiments, the data consistently supported the hypothesis that animacy drives visual processing independently of low-level features:

• Working Memory (Exp 1 & 2): Participants were significantly more accurate at detecting changes that altered animacy (e.g., rabbit $\to$ boot) compared to changes that preserved animacy (e.g., rabbit $\to$ dog), despite the boot and dog being the same image rotated.
  • Exp 1: $M = 3.7\%$ improvement, $p = 0.025$.
  • Exp 2 (no labeling phase): $M = 4.7\%$ improvement, $p < 0.01$.
• Visual Search (Exp 3 & 4): Participants found targets significantly faster in "mixed animacy" arrays (animate target among inanimate distractors, or vice versa) compared to "uniform animacy" arrays.
  • Exp 3: $M = 108\text{ms}$ faster, $p < 0.001$.
  • Exp 4 (no labeling phase): $M = 124\text{ms}$ faster, $p < 0.001$.
• Present/Absent Search (Exp 5 & 6): The search advantage for mixed animacy arrays replicated in a present/absent task.
  • Exp 5: $M = 94.4\text{ms}$ faster, $p < 0.001$.
  • Exp 6: $M = 78.5\text{ms}$ faster, $p < 0.001$.
• Orientation Control (Exp 7):
  • The search advantage was replicated for the original anagram stimuli ($M = 75.7\text{ms}$, $p < 0.001$).
  • The search advantage disappeared for the silhouetted stimuli ($M = 2.16\text{ms}$, $p = 0.86$).
  • The difference between the anagram effect and the silhouette effect was statistically significant ($p < 0.001$), confirming that orientation/aspect ratio alone cannot explain the results; the effect is driven by the perception of animacy.

5. Significance

• Cognitive Architecture: The results suggest that animacy is an ancestrally prioritized feature in human cognition, processed as a distinct high-level representation rather than a byproduct of low-level feature integration.
• Philosophical Implications: The findings address the "border between seeing and thinking" (Block, 2023), arguing that high-level representations (like animacy) can be disentangled from and processed independently of their low-level correlates.
• Methodological Advancement: By demonstrating that "identical" images can yield different cognitive outcomes based solely on interpretation, this work opens new avenues for studying other high-level concepts (e.g., emotion, social intent) using visual anagrams, moving beyond the limitations of naturalistic stimuli.

In conclusion, Boger and Firestone demonstrate that the human visual system is sensitive to animacy itself, providing robust evidence that high-level categorical perception operates beyond the constraints of lower-level visual features.