Image-generating AI, like the one on Yiaho, DALL-E, or Midjourney, excels at creating impressive images but often fails to draw realistic hands. Why? It comes down to:
- imperfect training data,
- the anatomical complexity of hands,
- generalization limitations,
- algorithmic prioritization,
- a lack of specific feedback.
But let’s look at this in more detail:
Why Can’t Image-Generating AI Create Hands?
In recent years, image-generating artificial intelligence has revolutionized creation. From futuristic landscapes to striking portraits, it sometimes seems to surpass human artists.
Yet, all it takes is zooming in on the hands to see its limitations: fused fingers, six phalanges instead of five, or deformed hands worthy of a horror film. Why this recurring failure on such a common detail? Behind these oddities lie technical and conceptual challenges that AI models still struggle to overcome.
Let’s explore five reasons that explain why hands remain AI’s Achilles’ heel.
1. Imperfect Training Data
AI models like Yiaho, Stable Diffusion, or Grok (in its visual applications) learn from massive databases containing millions of images. But hands pose a problem from the start: they’re not always well represented. In everyday photos, they can be blurry (a poorly framed selfie), hidden (behind an object), or captured from unusual angles (a twisted hand).
Even in art, illustrators often stylize hands in exaggerated or abstract ways, further muddying the waters. For example, an impressionist painting might simplify fingers into color blobs, while a manga exaggerates proportions. AI, which relies on this data to establish patterns, ends up with fragmented and unreliable “knowledge” about hand anatomy.
2. Underestimated Anatomical Complexity
Human hands are a marvel of nature. With 27 bones, over 30 muscles, and mobility that
allows gestures ranging from writing to playing piano, they surpass most other body parts in complexity.
A face, for example, follows a relatively fixed structure: two eyes, a nose, a mouth.
Hands, on the other hand, constantly change shape depending on their position, angle, or interaction with an object. Imagine an AI trying to render a hand holding a cup: it must understand the curve of the fingers, the shadow of the handle, and the texture of the skin—all at the same time.
Add variations between ages (child’s hands vs. wrinkled hands), sizes, or ethnicities, and you get a puzzle that current neural networks solve poorly, often producing twisted fingers or improbable joints.
Also read: Why Does ChatGPT Cut Off Its Sentences? Explanation and Solution
3. A Weakness in Generalization
Generative AI relies on Deep Learning, which excels at spotting patterns in training data. But when it comes to generalizing from rare or ambiguous examples, it shows its limits. Hands, with their countless poses (closed fist, crossed fingers, greeting gesture), require fine contextual understanding. If the AI has seen thousands of hands holding a pen, it might succeed in drawing one in that position.
But show it a hand playing with a ball or petting a cat, and it’s likely to “hallucinate”: too many fingers, not enough, or a shape that defies logic.
This problem is linked to the absence of true artificial general intelligence: current AI doesn’t “understand” hands, it imitates them based on what it has seen, and clumsily when the context changes.
4. Algorithmic Resource Prioritization
Creating an image via AI requires enormous resources: millions of calculations to transform a prompt into pixels. But these resources aren’t distributed equally. Algorithms are often optimized to prioritize main visual elements—an expressive face, detailed scenery—at the expense of secondary details like hands.
For example, if you request “a portrait of a woman in a garden,” the AI will emphasize her face and the flowers, relegating the hands to a quick approximation. This choice makes sense from a technical standpoint: a poorly drawn face ruins the image, while a strange hand sometimes goes unnoticed.
Moreover, rendering each finger precisely requires computing power that current models don’t always allocate, especially under time or energy constraints.
5. A Lack of Specific Feedback
AI systems improve through feedback loops, like Reinforcement Learning or human critiques. But errors on hands aren’t always flagged as a priority. When a user rejects an image, they’ll often say “it’s ugly” or “it’s not right,” without specifying “the hands are wrong.” Without this targeted feedback, the AI doesn’t know it needs to adjust this specific aspect.
Take Midjourney: if users simply rate images without emphasizing twisted fingers, the model continues to ignore this flaw. It would require dedicated training—for example, annotated datasets of correct hands or explicit critiques—for it to improve. In the meantime, hands remain a weak point due to lack of attention.
Will Image-Generating AI Improve?
These obstacles aren’t insurmountable. Researchers are already working on solutions: enriched databases with 3D hands, more powerful algorithms to handle complexity, and techniques like Transfer Learning to refine details.
AI could also benefit from Explainable AI to better understand its own errors. In a few years, twisted hands could become an amusing memory, a relic of generative AI’s clumsy beginnings. In the meantime, these imperfections remind us that even the most advanced technologies have their limits and sometimes, a strange charm!
