Synthetic image generation is genuinely impressive, and AI models produce convincing images across an enormous range of subjects. X-ray images turn out to be a category where they consistently fail, and the way they fail is more interesting than the fact that they do.
The short version: current models generate what an X-ray looks like without modeling what an X-ray measures. Every error below traces back to that distinction.
The difference between appearance and measurement
A photograph records light reflected from surfaces. An X-ray records how much radiation was absorbed along a path through an object, which means every value in the image corresponds to material density and thickness.
That is a physical relationship rather than a stylistic one. Dense material appears one way, low-density material another, and empty space another still, consistently, throughout the image. A generator trained on the appearance of X-ray images reproduces the aesthetic without any underlying model of density, which is why the results look plausible at a glance and fall apart under inspection.
The PCB
Comparing a generated circuit board against a real X-ray of one shows the problem clearly.
The most revealing error is the resistors, which appear as holes through the board. In an X-ray, white indicates air. A real resistor has low radiopacity, so it does appear light, but something is still there and the image shows it, along with the solder fillets at each end which are quite visible. The generator seems to have taken low density and rendered it as absence.
Capacitors show the inverse error. A capacitor’s dielectric material is highly opaque and appears distinctly dark in a real image. The generated board has nothing clearly identifiable as a capacitor at all.
Vias are wrong too. In a real X-ray, vias are typically open and read as such. The generated ones appear plugged with solder throughout.
Then there is the layer structure. A real multilayer board shows an orderly pattern, with traces running vertically on one layer, horizontally on the layer beneath, vertically again on the next. That orthogonal alternation is how boards are actually routed and it is unmistakable in a radiograph. The generated board reads as essentially single-layer, except for faint traces beneath one component that hint at an internal layer without following through.
Beyond the physics, the layout itself does not hold up. Traces run to nowhere, one component is simply missing, and the arrangement does not correspond to a circuit that could function.
The overall impression is of optical and X-ray conventions mixed together, as though the generator was drawing on both without knowing they are different things.
The iPhone
The phone is a more forgiving subject and the result is better, though not by much.
The large structures are approximately right. Three cameras in the top corner, a large battery occupying much of the body. But the battery is the wrong shape. Apple has used an L-shaped battery since the iPhone X, which is one of the most recognizable features of a phone radiograph, and the generated version misses it.
The rest of the interior is undifferentiated blocks. Most telling, the generator rendered visible PCB traces on the mainboard. At the magnification this image represents, those traces would be far too small to resolve. The generator produced detail that could not physically exist at that scale, which is a scale error rather than a drawing error.
The coffee maker and the toaster
These two share a tell, and it appears in both.
The generated coffee maker gets the general arrangement plausible, showing a pot and something resembling a heating element. But it renders plastic components as clearly visible structures. In a real X-ray, plastic is largely transparent, and what stands out is metal. The real comparison is a Keurig, where the aluminum pods inside are strikingly visible for exactly that reason.
The toaster shows the same category of error, with a heating element apparently floating somewhere it does not belong and a shaft-like structure that does not correspond to the pop-up mechanism a real toaster radiograph shows clearly.
And in both images, the generator placed a small block of generic circuitry in a corner. It appears to reason that an appliance is an electronic product, therefore some electronics should be visible, so a plausible-looking rectangle of components goes somewhere out of the way. It is a reasonable-looking answer arrived at without any model of what the device actually contains.
The component
A single component on a board is the simplest subject here and produced the weakest result.
The pads are rendered round. Real component pads are rectangular, and this is not a subtle distinction. Heavy trace-like structures run across the lead frame, possibly an attempt at wire bonds, though wire bonds are fine and these are not.
Most fundamentally, there is no die. The silicon is the entire reason the component exists and the reason anyone would X-ray it, and the generated image does not include one.
The ballistic plate
The generated plate has the right outline and an internally contradictory interior.
It renders as very light overall, consistent with a soft armor material like Kevlar. But it also has a dark, dense frame and a structured pattern of features resembling a cracked ceramic plate. Those are two different products, and the image is attempting to be both simultaneously.
The pattern
Across every example, the same failure mode repeats. The generated images capture the visual style of a radiograph, the grayscale palette, the ghostly transparency, the general sense of seeing inside something, without encoding the relationship between density and brightness that gives a radiograph its meaning.
That is why they look convincing to a general audience and obviously wrong to anyone who reads X-rays. Nothing in them is inconsistent with the aesthetic. Everything in them is inconsistent with the physics.
Not yet
This is a snapshot rather than a verdict. AI capability moves quickly, and the gap here is not obviously permanent.
What would close it is a model that understands X-ray images as measurements of material properties rather than as a visual style, and produces images where brightness means something. Until then, the generated images remain useful mainly as a demonstration of how much information a real radiograph actually contains.
