REAS

IN SILICO

In Silico extends Casey Reas’ sustained inquiry into how code can generate organic form. Since the early 2000s, Reas has written custom software systems to produce images from encoded rules, merging instruction-based art with the evolving capacities of computation. The In Silico series extends this practice using generative adversarial networks (GANs). Each work is generated from models trained exclusively on plant images that Reas foraged and then photographed through digital scanning and microscopy. The result is an ongoing body of work in which botanical form is reimagined through a medium whose mechanics are statistical, synthetic, and computational.

Within In Silico, Reas situates GANs in relation to photography. Just as a camera’s lenses, emulsions, and chemicals shape the field of possible photographs, a trained GAN defines the scope and style of its outputs. At the same time, artistic agency resides in the choices surrounding what to capture, how to “develop” the model through training, and how to select from (or intervene in) the resulting images. In this sense, In Silico is not about algorithmic novelty, but about learning a technological instrument—in this case, the GAN—deeply enough to perform with it.

The project’s botanical focus connects it to a longer history of artists and scientists who used emerging imaging technologies to catalog and interpret plant life. A key precedent is Anna Atkins’s Photographs of British Algae: Cyanotype Impressions (1843), one of the earliest photographic books. Her cyanotypes functioned as both scientific documents and aesthetic compositions, creating a novel alliance between botany and the new medium of photography. Reas’ datasets likewise emerge from fieldwork and careful collection, but here the transformation occurs not through chemistry and sunlight, but through neural networks and numerical training. In this way, In Silico positions machine learning as a successor in a lineage of technologies that extend human vision into new registers.

In Silico (2025)

The eponymous In Silico work represents both a culmination and a turning point within the larger series. Unlike the endlessly unfolding, time-based works that define much of Reas’ practice, In Silico (2025) consists of three large-format prints, each measuring 4 × 8 feet. These pieces do not move or recombine; instead, they condense the dynamics of generative systems into fixed images that confront the wall like an altarpiece.

In Silico, 2025.
Each print emerged through Reas’ navigation of the latent space of his trained GAN model, as he selected singular moments that balance organic form and computational distortion. In each composition, botanical structures loom and dissolve: veins unfurl like river systems, petals unravel into cellular lattices, and fragments of satellite topographies hover at the edge of recognition. The traces of neural-network training remain visible in each print’s glitches and textures, underscoring their synthetic origin. And yet, the works also insist on their continuity with analog images that capture the natural world.

In Silico extends the dialogue begun in Technical Images (2024), in which Reas paired GAN synthesis with nineteenth-century wet plate collodion photography. There, aluminum plates carried the chemical residue of collodion alongside machine-generated plant forms as a way to collapse the distance between photography’s earliest apparatus and machine learning’s statistical vision. In Silico (2025) carries this experiment forward, transposing the same hybrid logic into monumental prints.

By drawing together the GAN, the collodion plate, and the botanical archive, In Silico (2025) crystallizes the project’s central concern: how contemporary technologies extend, transform, and destabilize the ways in which nature is understood.

Earthly Delights

Earthly Delights expands the In Silico series from fixed images into animation. This ongoing series takes the form of nonlinear, generative software composed from frames produced during GAN training. Each work unfolds without a linear image sequence by recombining images at varying rhythms and tempos defined through Reas’ custom code. The works are indefinite in duration, continuously reshuffling their possibilities into new constellations.

The work’s title invokes Stan Brakhage’s 1981 film, The Garden of Earthly Delights, created from plants collected in Colorado and placed directly onto film stock. While Brakhage bypassed the camera and conventional filmmaking processes, Reas parallels this gesture while reimagining it through computational means. Here, plants are scanned rather than pressed, the apparatus is a GAN rather than a camera, and the outcome is a recombinant moving image untethered from the linearity of film.

VIDEO excerpt from Earthly Delights 1.1 SOFTWARE, 2019.

VIDEO excerpt from Earthly Delights 3.2 SOFTWARE, 2024.

As of 2025, three full series of Earthly Delights have been released. Each works with the same generative framework, yet produces distinct atmospheres. Some passages verge on the photographic, where blossoms or leaves briefly appear before dissolving into noise; others become fields of pure abstraction, with colors layered into painterly density. This balance of recognition and dissolution echoes both the fragility of natural forms, and the instability of digital processes.

The Studies for a Garden of Earthly Delights (2018) are closely related. These shorter works foreground the dataset itself, functioning as digital herbarium sheets or exploratory sketches. While less elaborate than the full series, they reveal the source material and the act of collection. Like preparatory drawings, they expose the iterative foundation of the larger works. Seen together, the studies and the full series emphasize recombination and variation—processes that define both botanical growth and artistic practice.

Study for a Garden of Earthly Delights, No. 1. 2018. HD video (color, silent). 03:00, loop.

Study for a Garden of Earthly Delights, No. 4. 2018. HD video (color, silent). 03:00, loop.

Technical Images

The works in Technical Images establish Reas’ dialogue with the material history of photography. Created in collaboration with artist Erika Weitz, they employ the nineteenth-century wet plate collodion process, one of photography’s earliest techniques. Rather than using aluminum plates to record more traditional subjects, such as landscapes or portraits, Reas uses them to fix GAN-generated botanical images. The resulting objects are hybrids: they bear the tactile presence and chemical irregularities of collodion, yet their imagery arises entirely from algorithmic synthesis.

The works explicitly reference philosopher and media theorist Vilém Flusser’s notion of the “technical image,” which describes how technical apparatuses (like cameras) mediate vision, and therefore condition the production of images. By embedding GAN imagery within collodion’s chemical substrate, Reas and Weitz staged an encounter between two paradigms: the capture of light on a sensitized plate, and the computational generation of form from data. In Technical Images, the artifacts of both—the swirls and imperfections of collodion, and the distortions and uncanny morphologies of GANs—are fused into single objects. In this convergence, the series situates machine learning not outside, but within the continuum of photographic processes, grounding the later In Silico works in a lineage that stretches back to photography’s origins.

Technical Image #2, 2024.
Collaboration with Erika Weitz.

Technical Image #5, 2024.
Collaboration with Erika Weitz.

RGB Technical Images

The RGB Technical Images grow directly out of the earlier Technical Images, extending that first experiment with wet plate collodion into a study of digital color. Each GAN-generated image was separated into its red, green, and blue channels, and each channel was rendered as an analog photograph on aluminum. Subverting the original color image and its RGB data, these three plates remain entirely black-and-white, preserving only the tonal information of their channels. Scanned and digitally recombined in a grayscale space, they form composites that resist the restoration of chromatic color.

By holding the channels in monochrome and layering them outside their intended roles, the works shift RGB from an invisible scaffold for color perception into a primary subject. The familiar structure of digital color is exposed but also unsettled, translated into a hybrid form where chemical artifacts—dust, streaks, and swirls—intermingle with the statistical distortions of GAN synthesis. The resulting images carry the density of multiple translations, reminding the viewer that photographic representation, whether chemical or digital, is always the outcome of layered processes and material constraints.

RGB Technical Image #1, 2024. Collaboration with Erika Weitz.

RGB Technical Image #3, 2024. Collaboration with Erika Weitz.

Amplified Technical Images

In contrast to the other works in this series, the Amplified Technical Images abandon chemical photography altogether. They work directly with GAN-generated imagery, digitally enlarging and intensifying chromatic shifts, pixel grids, and the uncanny distortions that emerge during training. Where the earlier works hybridized GAN synthesis with wet plate collodion, these pieces heighten the digital alone to foreground the raw materiality of computation.

Printed as dye sublimation on aluminum, the works remain in material dialogue with the collodion plates. The metallic surface recalls nineteenth-century photographic processes, but here it serves as a support for images born entirely from software. By amplifying rather than tempering GAN artifacts, these works invert Reas’ earlier strategy: instead of softening algorithmic traces through historical processes, they push them forward, insisting on the aesthetics of computation itself.

Amplified Technical Image #1, 2024.

Ampliifed Technical Image #2, 2024.

Datasets and Training

While the artistic focus of In Silico lies in the finished works, their creation depends on specific datasets and training processes. Reas has assembled two primary collections of plant imagery for use in this series: IS-1 and IS-2. The first, gathered in Colorado in 2018, includes scans of summer wildflowers and leaves from aspen forests and meadows, echoing the botanical material that inspired Brakhage. The second, created in 2024, combines scans of California native species with microscopic images of their internal structures and satellite views of the landscapes where these plants are found.

Colorado forest snapshots taken during the IS-1 dataset foraging, 2018.

Plant samples organized for scanning during the IS-1 dataset, 2018.

As Reas has built on the In Silico series in other bodies of work, these datasets have been used to train successive generations of GAN models: Deep Convolutional GANs (DCGANs), Progressive Growing GANs, and StyleGAN2. Each architectural shift altered the character of the outputs, from the blockier, unstable forms of early models to the heightened resolution and nuanced variation of later ones. As a particular instance of how a GAN model informs the work, In Silico (2025) draws on StyleGAN2 which, when trained on the second dataset, allowed the large-format prints to achieve their heightened clarity and depth.

In Reas’ work, the GAN models are treated not as ends in themselves, but as instruments. Like cameras, they establish the boundaries of what can be generated, yet artistic judgment operates at every stage: Reas is solely responsible for deciding what to collect, how to process it, how to guide the training, how to navigate the latent space, and how to curate the results. In this sense, the technical workflow is inseparable from the conceptual framework, situating the works within both the history of image-making apparatuses and the aesthetics of generative systems.

Continuities and Convergences

In Silico reveals a set of preoccupations that run throughout Reas’ practice. The series engages with systems, emergence, and transformation—concepts central to the 2004 Process works—but transposes them into the realm of machine learning. In Silico revisits the dialogue between art and science that underpinned early photography, while refracting it through contemporary neural networks. It also extends the experimental film lineage, specifically the work of Brakhage, by reimagining botanical material as the foundation for nonlinear, generative cinema.

Installation view of rgb technical images #1 and #3 in wet and saturated process at unit london, 2024.

Installation view of earthly delights 3.2 in Infinite Images: The Art of Algorithms at the toledo museum of art, 2025.

At the same time, the works converge on questions of medium and materiality. Wet plate collodion situates GAN imagery within photography’s chemical history, while dye sublimation prints on aluminum link digital images to the metallic substrates of nineteenth-century plates. The software’s infinite recombinations push cinema beyond the mechanical frame, while the monumental prints recall both scientific illustration and photographic murals.

Rather than a single aesthetic, In Silico presents a continuum of possibilities. Its images are at once natural and artificial, historical and futuristic, organic and computational. They acknowledge the imperfections of their apparatus—the glitches of GAN training, the streaks of collodion chemistry—and treat these not as flaws, but as defining qualities.

From Colorado wildflowers to Bay Area satellite images, from collodion plates to aluminum prints, from continuously generated video to monumental stills, the series spans centuries of image-making while articulating a distinctly contemporary vision. To see “in silico” is to perceive through computation itself—and to recognize how such vision reshapes our understanding of both nature and media.

Related Artworks
Browse the In Silico category on the REAS INDEX.

Further Exploration
Brakhage, Stan. The Garden of Earthly Delights. Film, 1981.

Flusser, Vilém. Towards a Philosophy of Photography. London: Reaktion Books, 2000.

Karras, Tero, Timo Aila, Samuli Laine, and Jaakko Lehtinen. “Progressive Growing of GANs for Improved Quality, Stability, and Variation.” International Conference on Learning Representations (ICLR), 2018.

Kent, Charlotte. “Casey Reas: In Silico.” Unit London, 2024.

Radford, Alec, Luke Metz, and Soumith Chintala. “Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks.” International Conference on Learning Representations (ICLR), 2016.

Reas, Casey. Making Pictures with Generative Adversarial Networks. Montreal: Anteism Books, 2019.

Schaaf, Larry J., with contributions by Joshua Chuang, Emily Walz, and Mike Ware. Sun Gardens: Cyanotypes by Anna Atkins. New York and Munich: Prestel Publishing, 2018.

Unit. “A Short History of Photography with Artists Casey Reas and Erika Weitz.” YouTube video, 3:19. Accessed August 20, 2025. https://www.youtube.com/watch?v=_XH76IJcQJU

Selected Exhibition History
Infinite Images: The Art of Algorithms. Toledo Museum of Art. Toledo, OH. 12 July – 30 November 2025. Curated by Julia Kaganskiy. Earthly Delights 3.2. [E-25-04]

Wet and Saturated Process. Unit London. London, UK. 24 April – 25 May 2024. Debut of new Technical Images and Earthly Delights 3.2. [S-28]

Interreality. Desmond Tower. Los Angeles, CA. 14 October – 25 November 2023. Earthly Delights 2.2. [E-23-10]

Co-Created: The Artist in the Age of Intelligent Machines. Burlington City Arts (BCA). Burlington, Vermont. 10 February – 6 May 2023. Earthly Delights 2.2. [E-23-03]

Illusionary Nature. Museum Sinclair-Haus. Bad Homburg vor der Höhe, Germany. 10 November 2019 – 2 February 2020. Curated by Ina Fuchs. Earthly Delights 1.1. [E-19-09]

Compressed Cinema. DAM. Berlin, Germany. 22 March – 4 May 2019. Debut of Earthly Delights. [S-21]