Diffuse Control by Beeple

Diffuse Control is a generative, interactive video sculpture created by Beeple. Rather than presenting a single fixed image or video, the artwork continually produces new visual combinations across twelve synchronized screens.

The work changes in response to source images selected by a curator, the parameters established by the artist, the behavior of its software, and interventions made by visitors.

No single participant completely determines what appears on the screens. That distribution of creative control is fundamental to the artwork.

Beeple is the artist who conceived and created Diffuse Control. Artificial intelligence is one component of a much larger artistic system that he designed.

The work reflects a series of human choices. Beeple determined the concept of the artwork, designed its physical form, selected and configured its technologies, established the rules governing how images behave, created the ways audiences can interact with it, and determined how those interactions become part of the work.

Curatorial choices introduce another layer of human agency. For MMoCA’s presentation, I selected works from the Museum’s collection and determined the particular groupings and relationships that enter the system.

Visitors introduce still another. Their actions influence what the system produces in real time, while the software interprets those interventions in ways that are not entirely predictable.

Diffuse Control therefore distributes creative agency among four principal forces: artist, curator, audience, and machine. None acts entirely alone, and none completely controls the result. That tension is fundamental to the artwork—and to its title.

The title describes the structure of the work itself.

Beeple establishes the system and its rules. As curator, I select the visual source material and establish relationships among those images. The software interprets and transforms that material. Visitors intervene in real time, sometimes individually and sometimes collectively.

No single participant completely determines the outcome.

Control is distributed—or diffused—among artist, curator, audience, and machine.

Knowing that an artwork uses oil paint, photography, video, or computer code tells us something about how it was made, but it does not tell us what the artwork ultimately is or means. I approach generative AI in much the same way.

Stable Diffusion is one of the technologies used to create Diffuse Control, but the technology alone is not the artwork.

The physical sculpture, Beeple’s artistic decisions, my curatorial decisions, the particular artworks selected, the behavior of the software, and the actions of visitors all determine the experience.

Diffuse Control is an experiential artwork. Its meaning cannot be separated entirely from encountering the sculpture, seeing the selected works transform over time, and experiencing how one’s own actions—and the actions of other visitors—affect what happens.

Understanding the technology can inform that encounter. I do not think it can substitute for the encounter itself.

Diffuse Control combines publicly available generative AI technology with proprietary software developed specifically for the artwork.

One of the technologies incorporated into the work is Stable Diffusion. Beeple uses this existing technology as one component within the larger system he designed for Diffuse Control, including software that allows selected source images to blend, transform, and move through a continuously evolving visual environment.

Diffuse Control is therefore not simply a prompt entered into a public image generator. It is an artistic system conceived and designed by Beeple in which AI, curatorial choices, source images, audience participation, and custom software continually interact.

Stable Diffusion is a widely available generative AI model capable of creating and transforming images based on patterns learned from a very large dataset of images and associated text.

The model was developed independently of Beeple and Diffuse Control. It has been made available for others to use and build upon, allowing artists, researchers, developers, and others to incorporate the technology into their own applications and systems.

Stable Diffusion is a model rather than a database of pictures that it searches and assembles each time an image is generated. During training, the model learned statistical relationships among images, words, visual characteristics, and concepts. Those learned relationships allow it to generate or transform images.

Beeple incorporates this existing model as one technological component within the much larger artistic system he designed for Diffuse Control.

The collection images selected for MMoCA’s presentation are introduced into Diffuse Control’s closed system and held locally as part of the artwork. They are not uploaded into a massive public training dataset or contributed to an outside AI platform for future model training.

Information identifying the source artworks accompanies each iteration and is preserved as part of the artwork’s documentation.

Artificial intelligence is used in several different ways within the artwork.

The primary image-generating system incorporates Stable Diffusion within the larger software environment developed for Diffuse Control. Other limited functions use additional technologies. For example, an OpenAI system assists in generating titles associated with the “Generate Extract” function.

These ancillary uses are separate from the system that transforms MMoCA’s collection images. The collection images remain within Diffuse Control’s closed image-generating environment and are not submitted to these outside services.

Visitors can influence the sculpture in real time through a kiosk control panel. 

Among the available controls are adjustments to color and movement through what the interface calls “latent space”—the range of visual possibilities available to the generative system.

Visitors can influence the artwork individually, but the system is also designed to respond to collective participation. Multiple visitors can interact with Diffuse Control simultaneously, sometimes cooperating and sometimes competing for control of what appears on the sculpture’s screens.

In generative AI, “latent space” is a mathematical way of describing relationships among possible images and visual characteristics.

A useful way to imagine it is as a vast landscape of visual possibilities.

Changing the latent-space control encourages the system to move through that landscape, sometimes remaining relatively close to recognizable aspects of the source image and sometimes moving toward more transformed or abstract results.

Diffuse Control is designed to respond not only to individual visitors but also to groups.

When multiple visitors interact with the artwork together, they can trigger special behaviors within the system.

Collab Mode rewards coordinated participation by unlocking a temporary visual event. Glitch Mode can emerge when many users compete for control at the same time, intentionally producing moments of instability and distortion.

These modes reinforce one of the artwork’s central ideas: control is never entirely held by a single person.

I selected the works through the curatorial process for MMoCA’s presentation of Diffuse Control.

As with any exhibition I curate, those choices consider visual relationships, historical context, artistic ideas, and the possibilities created by bringing different works into conversation with one another.

In Diffuse Control, however, those choices become part of the generative process itself. The particular works selected—and the relationships established among them—influence what the system is capable of producing.

The curator is therefore not simply supplying images to a machine. Curatorial selection is one of the human forces shaping the artwork.

Images introduced into MMoCA’s presentation of Diffuse Control are either in the public domain or used with permission from the appropriate copyright holder.

When permission is required, MMoCA obtains it from the appropriate rights holder.

Copyright can continue for many years after an artist’s death. During that period, copyright may be administered by an artist’s estate, foundation, family, designated representative, or another rights holder. Those rights holders have the legal authority to grant or deny permission for uses of copyrighted works.

This system serves an important practical purpose. Artists’ works do not disappear from cultural life when the artists themselves die. Estates and other rights holders routinely make decisions that allow works to be reproduced in books and exhibition catalogues, included in films and educational materials, licensed for new uses, and otherwise shared with audiences who may never have the opportunity to encounter the original object.

People may have different views about whether a deceased artist would personally have approved of a particular technology or artistic use that emerged after their lifetime. I think that is a legitimate subject for discussion. Copyright law addresses this circumstance by vesting authority during the copyright term in the artist’s legally authorized rights holder.

MMoCA works with the appropriate rights holder when permission is required.

An artwork in the public domain is no longer protected by copyright, either because its copyright has expired or because it was never eligible for copyright protection. It may therefore be reproduced, studied, published, adapted, and used to create new works without obtaining permission from a copyright holder.

The public domain is an essential part of the copyright system and reflects a longstanding principle that creative works can eventually become part of our shared cultural inheritance. Artists have always learned from, responded to, borrowed from, transformed, and reinterpreted the art that came before them.

Consider Leonardo da Vinci’s Mona Lisa. More than 500 years after Leonardo’s death, there is no artist or estate from whom permission could reasonably be sought each time the painting is reproduced or reinterpreted. Because the work is in the public domain, its image can circulate around the world—in scholarship and education, in books and films, and in countless subsequent works of art. Someone who may never have the opportunity to travel to Paris can nevertheless know, study, discuss, and respond creatively to the painting.

The same principle applies to new technologies. Public-domain works do not cease to be in the public domain because the tool used to study, reproduce, transform, or reinterpret them is digital or uses artificial intelligence.

For Diffuse Control, MMoCA distinguishes between works in the public domain, which may be used without permission, and works still protected by copyright. When copyright permission is required, MMoCA works with the appropriate rights holder to obtain it,

No. Stable Diffusion is an existing generative AI model that was developed independently of Diffuse Control and independently of MMoCA’s exhibition.

The images selected from MMoCA’s collection are introduced locally into Diffuse Control’s own system. They are not uploaded into the massive dataset originally used to train Stable Diffusion, nor are they being contributed to another public AI training dataset.

MMoCA’s participation in the exhibition does not add the selected collection images to the training dataset of the underlying Stable Diffusion model.

Yes, possibly.

Many of the images selected for Diffuse Control have circulated publicly online, in museum databases, publications, educational resources, and other digital sources. Given the enormous scale of datasets used to develop large generative AI models, it is possible—and in some cases may be likely—that images of these works have previously appeared in datasets used to train AI systems.

Stable Diffusion was trained before MMoCA selected the works for this exhibition. Their inclusion in Diffuse Control does not add them to Stable Diffusion’s existing training dataset.

As curator, I am interested in artists who engage directly with the technologies, ideas, conflicts, and conditions shaping contemporary life.

Artificial intelligence is already influencing how images are made, circulated, owned, interpreted, and valued. Artists are embracing it, questioning it, resisting it, criticizing it, and redefining it. I am interested in what artists have to say about that transformation and in the role art can play in helping the rest of us think about it.

Diffuse Control is especially compelling to me because it complicates the idea that an AI artwork is simply something “made by a machine.” Beeple conceived the work and designed its system. I make choices as curator about which artworks enter that system and the relationships among them. Visitors make choices that alter what happens in real time. The machine interprets and transforms those interventions in ways none of us can completely predict.

Artist, curator, audience, and machine all exercise agency, but none has complete control.

The development of large generative AI models raises serious questions about training datasets, copyright, consent, compensation, creative labor, and environmental impact. These questions remain the subject of significant legal, ethical, artistic, and public debate.

As curator, I am particularly interested in what artists have to say about the technologies, ideas, and conditions shaping the world around us. Artists have long adopted, questioned, appropriated, resisted, and transformed new technologies. Their work can reveal possibilities and problems that might otherwise remain abstract.

Beeple’s decision to work with Stable Diffusion is part of that artistic inquiry. Stable Diffusion is not itself Diffuse Control; it is one of the technologies the artist has chosen to work with. Beeple has constructed a much larger artistic system around it—one that distributes creative agency among artist, curator, audience, and machine and asks us to consider where authorship and control reside when all four participate in making an image.

Artists hold widely differing views about generative AI.

Some embrace it as a new creative medium. Some use it critically or experimentally. Others object to aspects of its development or use. Artists and others have raised important questions about the sources used to train large AI models, consent, copyright, compensation, creative labor, environmental impact, and the ways automated image generation may alter cultural production.

Those disagreements are themselves part of the cultural landscape in which Diffuse Control exists.

As curator, I am interested in creating space for those differing perspectives rather than resolving them in favor of a single position.

My role is to facilitate an informed encounter with an artist and the questions his work raises. Visitors may ultimately embrace the work, reject it, or find themselves somewhere more complicated in between.

Artificial intelligence requires computing power, and computing infrastructure uses electricity and other resources, including water.

The broader environmental impact of AI—including energy consumption associated with model development, computing, data centers, and cooling—is an important and evolving area of public discussion.

Environmental impact belongs alongside questions of authorship, copyright, consent, and creative labor in an informed discussion of emerging AI technologies. It is one of the many questions that can appropriately accompany an encounter with Diffuse Control.

Contemporary art does not require consensus.

Visitors may arrive at very different conclusions about Diffuse Control, generative AI, Beeple’s choices, or my curatorial decisions. I think disagreement is entirely compatible with a meaningful encounter with art.

What I hope to provide as curator is accurate information, meaningful context, and an opportunity to encounter the artwork itself. I invite visitors to see Diffuse Control, interact with it, ask questions, challenge it, and form their own conclusions.

MMoCA welcomes criticism, disagreement, and peaceful protest. Museums should be places where difficult ideas can be examined and where people with differing perspectives can engage seriously with art and with one another.

Since announcing Diffuse Control, however, MMoCA and members of our staff have also received a small number of messages containing threatening language, including messages that threaten harm or illegal acts. MMoCA unequivocally rejects such behavior.

There is an essential difference between forceful criticism—even criticism with which the curator, artist, Museum staff, or others may strongly disagree—and language that threatens the safety of museum staff, artists, visitors, or others. Threats do not advance the important ethical and artistic questions surrounding generative AI, and they have no place in a productive civic conversation.

We ask our community to reject threatening language and behavior regardless of one’s position on this exhibition.

MMoCA remains committed to vigorous, open, and respectful exchange. As curator, I hope that exchange begins with curiosity: with looking closely, asking questions, listening to artists and one another, and remaining open to the possibility that an encounter with art might complicate what we thought we knew.