Before 1940, records were sold in plain kraft-paper sleeves — brown, unmarked, functionally indistinguishable from the envelope an electricity bill arrives in. The change came from a young designer named Alex Steinweiss, hired by Columbia Records as its first art director, who argued that the package could do something the disc could not: make a stranger curious. His illustrated sleeves sold so conspicuously well that the industry reorganised itself around them within a decade. An entire visual discipline was founded, essentially, on a packaging argument.

That origin story is worth remembering now, because the discipline is passing through its second structural change — and the argument being made this time sounds remarkably similar.

The gallery that fit inside a record crate

For roughly thirty years, the twelve-inch square was one of the most widely circulated formats in visual culture. It was not shown in institutions; it was carried home in paper bags, propped against walls, and studied at length by people who had no particular interest in graphic design. Few exhibition formats have ever had that reach.

The work made for it holds up accordingly. Reid Miles, art director at Blue Note Records from the mid-1950s, produced several hundred covers built from restricted colour, aggressive cropping and typography treated as the primary image rather than a caption — frequently pairing his layouts with Francis Wolff’s session photographs. The Hipgnosis studio, founded in 1968, spent the following decade constructing elaborate photographic scenarios in-camera, often at considerable expense and inconvenience, on the grounds that a viewer can tell the difference between a real event and a composited one. Robert Freeman’s half-lit portrait for the Beatles’ second album did more to fix that group’s image than any promotional campaign.

None of these are records of good taste alone. They are records of a specific production economy: a label with a budget, a commissioned specialist, a physical object large enough to reward close looking, and a lead time measured in weeks.

The thumbnail broke the composition

That economy dissolved twice. The first collapse was physical. A cover designed to be read at thirty centimetres is now first encountered, in most cases, at a few hundred pixels — a tile in a streaming grid, a lock-screen square, a circle on a smart speaker display. Detail that once repaid attention became noise. The formal response has been consistent and slightly depressing: heavier type, higher contrast, fewer elements, centred subjects. Legibility at small scale is a genuine constraint, but it is also a flattening one, and it applies whether the cover was made by a studio or a teenager.

The second collapse was economic. Most music released today is self-published. An artist uploading to distribution platforms needs cover art before the release can be scheduled, and the historical options were to commission a designer, license a stock image, or produce something themselves in software they had never learned. For a release with no advance and no label, the first option was rarely available, which is why so much independent release art of the past fifteen years defaulted to the same handful of solutions.

What generative models actually changed

Image models altered one variable in that equation, and it is worth being precise about which one. They did not make artists better at composition. What they changed is the cost of iteration.

Under a commission, an artist gets two or three concepts and picks one, because each additional round costs money and days. With a generative tool, twenty variations of a single idea can be produced in an hour and compared side by side — which is a different creative act altogether. It is closer to editing than to art directing. The skill that gets rewarded shifts from making the image to recognising which of twenty images is the right one, a judgement musicians turn out to be reasonably good at when it concerns their own work.

The practical tooling has followed that shift. A current album cover generator is expected to take a reference image and remix it rather than only work from a blank prompt, and to output the same concept across several aspect ratios — the square for streaming, a wide crop for banners and video platforms, a vertical version for short-form clips. That last requirement is genuinely new. Steinweiss designed one object. A release today needs the same idea to survive four crops without losing its centre of gravity.

Some of this tooling is now built directly into the platforms artists use to produce the music itself, so that artwork is generated in the same session as the track rather than outsourced afterwards. Lacuna is one of several services that took this route, treating the cover as part of the release rather than a separate procurement problem. Whether integration improves the results is arguable; that it removes a bottleneck is not.

What did not change

The concept still has to exist before anything is generated, and no model supplies it. This is the part most often misread in arguments about these tools.

Consider what Hipgnosis was actually selling. Not photographic technique — competent photographers were plentiful. The studio was selling the decision to photograph a man shaking hands with a man on fire, and the judgement to leave the image unexplained. The execution was expensive, but the execution was not the scarce part. Had those designers been handed a generative model, they would have produced the same body of work faster, and the images would still have been theirs, because the idea preceded the tool.

The inverse also holds. An artist with no visual position generates twenty images and cannot tell which is right, because there is no brief against which to judge them. Lowering the cost of production raises the value of editorial judgement; it does not distribute that judgement more evenly. Anyone who has looked at a large volume of recent independent release art will have noticed both effects at once — a visible rise in technical polish alongside a distinct sameness of imagination.

The authorship question is not settled, and probably should not be

For collectors, archivists and institutions, the harder issue is attribution. A Reid Miles cover has an author, a date, and a documented working relationship with a photographer and a label. A generated cover has a prompt, a model, a set of training data of contested provenance, and a person who selected the final frame. Those are not equivalent, and the catalogue conventions to describe the difference do not yet exist in any settled form.

This is a live problem rather than a hypothetical one. Sleeve design has been collected, exhibited and written about as a discipline for decades, on the understanding that a named designer made deliberate choices. Any serious history of the current period will need to record how an image was made, not only what it depicts — and that record is being created right now, largely without anyone keeping it.

A packaging argument, again

The instinct to treat cheaper image production as a decline in the form is understandable and probably wrong, or at least premature. The twelve-inch cover was itself a commercial device before it was anything else, dismissed as marketing by people who took music seriously, and it produced work that hangs in museums.
What determines whether the current period produces anything comparable is not the sophistication of the tools. It is whether artists using them have a position worth expressing in an image, and whether anyone is documenting the results carefully enough for the good work to be found later. Steinweiss did not have better printing technology than his competitors. He simply understood, before they did, that the sleeve was a place where something could be said.