In 1975, Cambridge’s Museum of Classical Archaeology painted a cast of the Peplos Kore in vivid colours. The point was to unsettle the familiar image of ancient Greek sculpture as white marble. The museum is equally clear that it does not know whether the original statue looked exactly like its reconstruction. The cast could challenge a misconception without being a recovered original. Cambridge’s account makes that distinction unusually candid.
The uncertainty reaches beyond shades of paint. Cambridge describes an alternative reconstruction whose interpretation of the clothing suggests a goddess. The Acropolis Museum, which holds the original, also discusses a possible goddess, perhaps Artemis, drawing on the clothing and the position of the hands. A restored garment or object placed in a missing hand can participate in an argument about who the figure is.
That is the problem I would want an AI restoration system to expose. Completing an image can require decisions about its meaning. Keeping the damaged image untouched does not spare us interpretation either: the loss of paint helped make white marble seem like the intended appearance.
Which image is being recovered?
Two recent research designs make the choice of objective concrete. In SAG-MR, published on August 21, Fubo Wang and colleagues use structure-guided diffusion to restore mural images. Their reference images are relatively intact pictures of murals. The paired training and numerical evaluation largely use damage added to those pictures, such as simulated cracks or fading. The authors explicitly distinguish these reference targets from the murals’ original historical appearance; genuinely damaged murals lack exact clean counterparts and are assessed qualitatively.
The distinction changes what a successful output would mean. Recovering the reference picture would undo a particular degradation of an image. It would not establish which pigments were present centuries earlier or whether an earlier conservation intervention belongs in a depiction of the original. The target supplies an answer to “which state?” before the model begins. Calling the process restoration does not supply that answer for a museum.
GeAR, revised on July 26, makes a different tradeoff explicit. Qinyu Zhang and colleagues start with a painting and construct a three-dimensional scene. First they translate its depiction toward more coherent geometry and illumination cues. Then they restore painterly appearance across views while constraining changes to the reconstructed geometry. The authors recognise that features troublesome for physical reconstruction can be essential to the artwork’s style.
That sequence is revealing. A painting need not describe a scene that a camera could have photographed. Turning it into a navigable space therefore calls for assumptions about how depicted forms continue beyond the original view. Restoring the brushwork afterwards can make that chosen space resemble the painting, but cannot establish that it was the space the painter intended. GeAR has no observed alternative viewpoints against which to establish that claim. Its rendering is an interpretation with a useful new affordance.
The version that leaves the archive
I do not think this makes reconstruction a mistake. Cambridge’s coloured cast has a defensible purpose precisely because bare marble can give visitors the wrong impression. A reconstruction can make a supported proposition legible while leaving some of its particulars unresolved. Demanding certainty about every detail would prevent that kind of explanation.
There is an established language for this. The London Charter asks digital heritage visualisations to make clear whether they represent an existing state, an evidence-based restoration or a hypothetical reconstruction. It also asks that significant dependencies between hypotheses be identifiable. AI did not invent this obligation. It gives software more of the choices through which a visual argument is assembled.
I would carry that distinction into what a system publishes. Suppose a museum retains the original photograph and a careful account of alternative interpretations, but exports only one seamless reconstruction for use elsewhere. If that picture circulates without the account, its chosen garment and restored attributes still travel with it. The qualification does not. This is a possible consequence of the publishing arrangement, not a claim that these studies measured such reuse.
Preserving the files has therefore left a consequential decision unresolved: which disagreement will the next viewer actually encounter? An archive can be meticulous while its default public image offers a single answer. A caption saying “AI reconstruction” identifies a process, but tells me little about whether the uncertain part is a colour’s intensity or an attribute that supports identifying a goddess.
I would give alternatives that change identity or historical interpretation a place in the first view, with the surviving object available alongside them. I would not fill the display with arbitrary generations. The alternatives need reasons rooted in the object and its scholarship. Less consequential variations could remain in the accompanying account.
That leaves a curatorial judgment which better image generation cannot make on its own. How much visible disagreement can a display carry before it obscures the supported idea it was meant to explain? Cambridge’s cast makes the tradeoff real: a forceful correction to the white-marble picture necessarily takes a more specific visual form. The task is to choose that form without allowing its persuasive detail to decide an unsettled historical question on the viewer’s behalf.