Artificial Intelligence, Film Restoration, and Early Cinema: Notes from AMIA 2022

By Hugo Ljungbäck

Over the past three years, popular interest in early film and moving image culture has seen quite a revival on social media, and YouTube especially, as it has become the home of hundreds, if not thousands, of “artificial intelligence restorations,” as they have become known: early actualities, travelogues, scenics, and other films have been upscaled to 4K, stabilized, smoothed out at 60 frames per second, sometimes cropped to widescreen format, and almost always colorized. Often sound is also added to make these historical images—of Paris, New York, Tokyo, Berlin, or Amsterdam—come alive in the present. The results—often uncanny—have generated significant interest in these early films, not just from viewers, who frequently post awe-struck comments about the eerie present-ness of the past, but also news outlets, which have helped share and circulate these clips with their catchy headlines: “AI Magic Makes Century-Old Films Look New,” or “‘Time Travel’ to the 1890s in AI-remastered Silent Movies That Look Like HD Video,” or—my personal favorite—“Watch this amazing AI-enhanced montage of Lumière films—historians will hate it though.”


Two comments taken from a “restored” video of New York in the early 1900s.


While most archivists and early film historians who were alerted to this trend may indeed have originally wanted to dismiss these videos as part of a short-lived, contemporary fad, the significant viewership they have developed and the attention they continue to be paid suggest that it might be time to start taking them seriously. And so, inspired by my own fascination, bemusement, and sometimes frustration with these clips, I proposed an open forum on the topic at the recent Association of Moving Image Archivists Conference in Pittsburgh. Not sure what to expect, I was pleasantly surprised by the significant interest in this conversation—to the point that we ran out of seating. Over sixty archivists, film restorers, technical vendors, rightsholders, academics, and students participated in the hour-long, informal discussion, which turned into a passionate debate about technology, ethics, and the future of our field. Here, I give a brief account of the main topics of concern that emerged and introduce some of the many more questions that arose, which are ripe for further discussion and consideration.

The open forum was an attempt to begin to grapple with these “restoration projects” within our field. From the hobbyist restorers’ point-of-view, these projects are serious and sincere attempts at restoring or improving the quality of the original films and making these historical images widely available and accessible to contemporary audiences; but from professional archivists’ point-of-view, they tend to be seen as creative practices that distort the historical value of the original images as well as misrepresent the work of professional restoration. As the hobbyists themselves are actively engaging in these discourses by making commentary and how-to videos where they reveal their processes and show the AI and neural network tools at work, how can we address these clips as both potential assets and liabilities in larger popular conversations about the preservation, restoration, and conservation of our moving image culture? Throughout the conversation, three general areas of concern emerged: the distortion of the image as historical index, the muddled meaning of restoration within our own field, and the potential for future collaborations with and use of artificial intelligence tools in professional restoration contexts.

The most vocal critics of hobbyist AI restorations have been those archivists and historians concerned with the ways in which the tools employed in AI restoration alter, transform, and divorce the image from its historical contexts. Because these processes rely on additive guesswork, creating in-between frames or elements of images that never existed to begin with, the image can no longer be trusted as a representation of the past. Many have already voiced their (often strong) opinions on Peter Jackson’s They Shall Not Grow Old, a 2018 film commissioned by the Imperial War Museum to commemorate the centenary of World War I, which used upscaled, colorized, and digitally manipulated footage from the IWM Archives to bring the soldiers back to life. The film has become a locus for archivists and historians concerned with AI technologies’ distortion of history, and EYE Film Museum is only one of several institutions that have addressed the controversial film through their public outreach. As “deepfakes” have already complicated the relationship between reality, truth, and the contemporary image, the prospect of having these tools unleashed on the mass-digitized historical archive without prescribed professional ethics and standards represents a slippery slope and should be cause for concern for a “downward spiral” of the historical record.

At the same time, some hobbyists agree that Jackson went too far in his manipulation of the original material and distance themselves from the kind of creative intervention and freedom he exercised over the IWM Archive footage. They proclaim that their own work is really only about enhancing the image that is already there, and that they only do it to make it easier and more pleasant for the viewer to look at—and because they enjoy it. And, as many early films have typically only been available online in heavily degraded condition, as low-res files digitized from VHS tapes, themselves telecined from bootleg 16mm prints, can we fault non-expert viewers and hobbyists for thinking that their AI enhancements look better? Rather, what can we do through our own work to make professional restorations of early films more accessible to the average viewer without institutional credentials and access to specialized finding aids? And how can we make these different versions—original, restored, and AI-enhanced—coexist, but contextualized?

For a lot of viewers, the enhanced videos on YouTube might be the first time they hear about film “restoration,” or the first time they encounter early films in general. As some AMIA attendees noted, the hobbyists, unlike their cash-strapped, professional counterparts, have the bandwidth and time on their hands to market themselves, to engage with their audiences, and to “control the narrative” about restoration in a way many archives cannot. As the mass digitization of media history in many ways already makes the work of archives and archivists invisible, these misconceptions about restoration and preservation are left unchallenged. But as several participants also remarked, the definition of “restoration” is already a heated, controversial, decades-old debate among professional archivists and restorers. Does restoration mean simply cleaning and repairing damaged film elements? Bringing back the original camera negatives and recutting them? Rescanning a first-generation film print at a higher resolution? How should we differentiate between film restoration (using photochemical means), digital restoration (using digital tools), and intellectual property restoration (returning a film to a director’s original creative vision, regardless of process or medium), or between restoring the image as content or the film element as its physical carrier?

Whatever we mean by “restoration,” some suggested that the term in general is too opaque, and that archivists and restorers need to be more transparent with scholars and the public alike about their interventions. “What did we do to this film, and why?” If people had a better understanding of what restorers did—what elements were used, what choices were made, and who made them—they could better appreciate their work, differentiate between the different versions of films they produce, and understand the decisions involved. Should you remove dust particles and scratches from a film scan? Should you artificially fix tears in a film strip, even if they were part of how the original film was first experienced over a century ago? As several restorers reminded us, “you don’t just push a button.” Someone must provide the algorithms with instructions and feedback on how extensively to intervene, whether based on a client’s requests or the restorers’ artistic sensibility. How do we decide what kinds of alterations are acceptable, and to what degree? Where do we draw the line between restoration and distortion? These are just some of the many questions we face in the next decade, as artificial intelligence tools, whether used by hobbyists or professionals, become ubiquitous in the field.

Either way, these AI restorations—or remasters, remixes, remediations, enhancements, or whatever they end up being called—represent a “teachable moment,” an opportunity to set the record, to define our vocabulary and standardize guidelines, but without gatekeeping and closing off the possibilities of future technologies and practices. As AI technologies have already been part of professional restoration toolkits for two decades, AMIA could play a role in leading these conversations by inviting other stakeholders—including hobbyists and other content producers—to the table. And early cinema and silent film scholars should be vying for a seat, as the widespread availability of these videos on YouTube is changing the way general audiences understand and interact with early film culture. While the most popular subjects of hobbyist restoration have so far been nonfiction films, whose “documentary” footage of people and places is ideal for historical time travel, several fiction films have recently been subjected to the same processes, ranging from Edwin Porter’s The Great Train Robbery (1903) and Lois Weber’s Shoes (1915) to Vsevolod Pudovkin and Nikolai Shpikovsky’s Chess Fever (1925) and beyond. Whether we like it or not, audiences are demanding more, and we’d be remiss to dismiss their engagement with these new incarnations of early films, and the old myths that accompany them.

A comment from a “restoration” of Arrival of a Train at La Ciotat, commenting on the immersive quality of the “restored” image and engaging with one of the founding myths of early cinema spectatorship.


At the end of the day, as archivists and historians, we want people to care about our moving image heritage, and if AI enhancements become an entryway—a gateway drug, if you will—that piques people’s curiosities, let’s harness this interest in archival film for our shared preservation and pedagogical missions. However foreign artificial intelligence may seem from our own discipline, no matter how distant these hobbyists may appear from the professionals, and whatever we may personally feel or think about AI restoration—fascination, bemusement, frustration, or a combination of all—we can all find common ground in our appreciation and love for the wonder of early cinema. And that’s plenty.