Why AI Food Looks Like That

Worms, holes, and cracks reveal the technical weaknesses of image generators.

If you've spent any time scrolling through AI-generated food images—whether on social media, marketing campaigns, or creative portfolios—you've likely noticed something unsettling. The salads might look perfect from afar, but zoom in, and you'll spot strange worms wriggling in the lettuce. The cakes, though beautifully frosted, may hide subtle cracks or gaps that defy the laws of physics. These quirks aren't random glitches; they're windows into the fundamental limitations of today's image generation models.

The Anatomy of an AI-Generated Meal

At the heart of these artifacts lies how modern generative models, such as diffusion-based systems (the technology behind tools like Midjourney and DALL·E), learn to create images. These models are trained on vast datasets of real-world photos, learning statistical patterns of shapes, colors, and textures. When generating a new image, they reconstruct it from noise, adding detail iteratively based on learned probabilities.

However, this process doesn't truly 'understand' food. A model might know that a strawberry is red and speckled, but it doesn't grasp that a strawberry's seeds are mechanically attached to its surface, not floating beneath the skin. Similarly, it may render a slice of bread as a homogenous mass, missing the fibrous structure that gives it coherent form. The result: images that look tantalizingly real at a glance but fall apart under closer inspection.

Why Worms, Holes, and Cracks Appear

The specific flaws—worms (thin, elongated structures), holes (unexpected voids), and cracks (discontinuous surfaces)—stem from the model's inability to maintain consistent object permanence and physical plausibility. In AI food imagery, these often appear because the model struggles with objects that have complex internal structures or repetitive, irregular patterns. For example:

• Worms in produce: Leafy greens or grains have high-frequency textures that confuse the model, leading to over-exaggerated lines that resemble living organisms.
• Holes in baked goods: Breads and pastries rely on precise air pockets and yeast action; AI models might over-feature these due to inconsistent training data, creating unrealistic voids.
• Cracks in surfaces: The interplay of crust, glaze, or toppings requires a depth of understanding (literally) that 2D-trained models lack, projecting shadows or breaks where none should exist.

These artifacts are not just aesthetic annoyances—they reveal a deeper truth: AI lacks commonsense reasoning about the physical world. It doesn't know that a tomato has skin, that cheese melts uniformly, or that a sandwich layer's structural integrity depends on moisture content.

Progress and Persistent Challenges in 2026

As of 2026, substantial strides have been made. Leading models now incorporate 3D-aware training and physics-based simulations to reduce such flaws. Some platforms allow users to generate food images with remarkable fidelity, especially when prompted with specific descriptors. Yet, challenges remain. The push for faster generation, lower computational costs, and higher resolution often reintroduces the same weaknesses. Moreover, the growing trend of personalized and niche AI imagery—like recreating nostalgic family recipes—demands a level of detail that current models still struggle to achieve.

Researchers are exploring hybrid approaches that combine neural networks with explicit modeling of material properties, but these are in experimental stages. Meanwhile, casual users can mitigate artifacts by using detailed, clear prompts (e.g., 'a rustic sourdough loaf with visible scoring and a glossy crust, photographed from a 45-degree angle with soft natural light') or by post-processing with image editing tools to fix obvious flaws.

Why It Matters

Understanding these quirks is more than a curiosity for foodies or tech enthusiasts. As AI-generated images become ubiquitous in advertising, cookbooks, and restaurant menus, it's essential to recognize their synthetic nature. For professionals, this knowledge is a reminder to verify AI output against reality, while for consumers, it offers a lens to see through dazzling but often deceptive portraits of what we eat.

So, the next time you encounter an AI-generated pizza that looks too perfect—or too creepy—remember: those worms, holes, and cracks aren't failures of imagination. They're the fingerprints of a technology still learning what a meal truly is.

via The Verge AI