The Mathematics of Disgust: Why AI Art Struggles with Clustered Holes
The 'Cursed' Image Phenomenon
If you have spent any significant amount of time exploring generative AI tools like Midjourney, DALL-E, or Stable Diffusion, you have likely encountered them—the 'cursed' images. You might be generating a portrait, a landscape, or a piece of abstract art, only to find the result riddled with strange, clustered bumps, uneven textures, or an unsettling array of holes. For many, these images are more than just aesthetically displeasing; they trigger a visceral, physical response known as trypophobia.
Trypophobia, while not currently classified as a specific phobia in the DSM-5, is a well-documented psychological aversion to clusters of small holes or bumps. But why does AI, a tool built on cold, hard logic, seem to have such a penchant for generating images that trigger this specific human fear? Is there a hidden link between AI detection abilities and this biological response? Let's dive into the technical reality behind these patterns.
The Intersection of Math and Phobia
To understand why AI generates these images, we must first separate the human experience from the machine process. Humans experience trypophobia as a fear or disgust response, likely rooted in evolutionary adaptations that helped our ancestors avoid dangerous animals, parasites, or decaying matter. When we see a cluster of holes, our brain instinctively flags it as a potential threat or a source of contamination.
AI, on the other hand, has no feelings, no fears, and no concept of 'disgust.' When a computer vision model like a Vision Transformer or a CNN (Convolutional Neural Network) looks at an image, it is not 'seeing' in the human sense. It is performing complex mathematical classification. It identifies edges, contrast, color gradients, and textures.
When these models are trained on massive datasets—containing billions of images—they learn to recognize patterns. As computer vision researchers have noted, AI models often gravitate toward 'trypophobic' textures because they are statistically dense in information. These patterns are visually 'loud'—they contain a high number of edges and high-contrast regions per square inch. During the training process, the model learns that these dense, complex textures are a reliable way to fill space and create 'detail,' which is often rewarded by the training objective.
Why AI 'Loves' These Creepy Patterns
There is a growing theory among online communities that AI models 'learn' to include these patterns because they are common in high-resolution, noisy, or distorted images found in training data. When a model is asked to generate a complex texture—like skin, bark, or fabric—it has to make a probabilistic guess about what that texture looks like.
Because of the way these models prioritize high-frequency patterns to achieve sharpness, they often default to repetitive, clustered structures. This is a technical coincidence, not a design choice. The AI is essentially taking a shortcut. It identifies that a 'complex surface' usually involves a lot of repeating contrast, and it generates that contrast in the most mathematically efficient way it knows: through repetitive clusters of dots or holes.
It is important to debunk the myth that AI is intentionally generating these images to 'creep out' users. This is a classic case of anthropomorphism—assigning human intent to a non-human system. The AI is not trying to be scary; it is simply doing its job of pattern generation, and sometimes, the output happens to overlap with biological triggers for human disgust.
How to Tame the Beast: Practical Tips for Better Generations
If you are sensitive to these patterns, or if you simply want to avoid the 'cursed' look in your projects, you do not have to accept these results as inevitable. Because these patterns are a byproduct of how the model interprets textures, you can use prompt engineering to steer the AI away from them.
Here are a few actionable strategies to clean up your AI generations:
Use Targeted Negative Prompts: Most advanced AI tools allow for 'negative prompts' (what you want to exclude). Explicitly list terms that trigger your discomfort. Try adding phrases like: 'clustered holes', 'trypophobia', 'excessive pores', 'distorted textures', 'repetitive bumps', 'noisy artifacts', 'irregular skin texture'.
Adjust Stylistic Influence: Often, models are pushed toward hyper-realism, which can sometimes lead to the AI 'hallucinating' excessive detail in an attempt to look realistic. If you are getting trypophobic patterns, try adjusting the 'style' or 'artistic influence' parameters. Moving toward a more 'painterly', 'smooth', or 'stylized' aesthetic often forces the model to simplify textures, reducing the density of those high-frequency patterns.
Post-Processing for Texture: If an otherwise perfect image is ruined by a small area of clustered holes, use in-painting or generative fill tools. You can mask out the offending area and prompt the AI to 'smooth skin' or 'fill with flat texture,' effectively removing the cluster without discarding the entire image.
Final Thoughts
The link between AI and trypophobia is not a psychological connection, but a technical one. It is a fascinating, if sometimes unsettling, reminder of how differently machines and humans process the world. While we see danger and disgust in clustered patterns, AI sees only information density and edge contrast.
By understanding that these images are merely a byproduct of mathematical optimization, we can move past the conspiracy theories and focus on the practical tools we have to control the output. The next time you see an AI-generated image that makes your skin crawl, remember: it is just math, and with the right prompts, you have the power to fix it.
