RaiSR RuiNS (2016–2017) — by Frederik De Wilde
RaiSR RuiNS began as a technical question with an aesthetic consequence: what would happen if you dramatically enlarged the fragile, low-resolution outputs of early generative networks?
RaiSR RuiNS radicalizes the residual errors of early generative systems. By upscaling fragile network outputs and subjecting them to learned sharpening, the series converts algorithmic artifacts into architectural matter. Where later work would pursue seamless photorealism, De Wilde’s images insist on the readability of process: here, resolution is excavation; upscaling becomes ruin; the artifact is the image.
The project took place when most GAN images still arrived as compact, pixel-bound thumbnails — the sort of images that read well at web scale but conceal the models’ internal scaffolding.
De Wilde’s procedure was simple and deliberate: take images produced in 2016 at roughly 227 × 227 pixels and scale them up — in this case to approximately 5,448 × 5,448 pixels (about a 24× magnification) by early 2018.
At those dimensions, computational structures that are normally invisible become the work’s subject. The large, repeating checkerboard motifs found across the series are not JPEG glitches or accidental compression: they are structural artifacts that arise from the upsampling operations used in early convolutional generative architectures.
These artifacts were carefully documented in a 2016 essay published by Distill, which explains how uneven overlap in transposed-convolution (so-called “deconvolution”) layers produces visible checkerboarding when feature maps are scaled.
In RaiSR RuiNS, that side effect is read as material: grids suggest masonry, interpolation seams read like fissures, and the image surface behaves like an archaeological cross-section through algorithmic construction. Generative engines and hybrid neurons The project’s generator strategy is rooted in techniques contemporary to Plug & Play Generative Networks (PPGN) (2016), a method that allowed images to be created conditionally by optimizing in latent space so as to activate specified neurons in a pretrained classifier. Unlike one-shot GAN sampling, PPGN performs iterative refinement: an image is nudged again and again until particular internal activations respond in the desired way. This workflow appealed to De Wilde’s interest in what was termed “hybrid neurons”.
For RaiSR RuiNS, De Wilde collaborated with pioneering machine-learning scientist and collaborator G. Goh. In the first phase a large image corpus trains classifiers and yields the neuron space; in the second, images are synthesized by blending and optimizing across those neuron activations. The system treats activation patterns not as invisible statistics but as sculptural medium — a way of “carving” images out of internal networks.
In parallel to the iterative generation, De Wilde experimented with early super-resolution techniques. RAISR — developed by Y. Romano, J. Isidoro and P. Milanfar and later presented via Google Research — takes a different approach to enlargement than naïve bicubic interpolation. RAISR trains a bank of directional filters on pairs of low- and high-resolution patches, then applies optimized sharpening in a single, extremely efficient pass. But when RAISR’s sharpening is applied to GAN-made imagery, the result is paradoxical: rather than restoring a lost clarity, the learned filters amplify the generator’s seams and structural idiosyncrasies. Sharpening becomes intensification — a process that clarifies not what the image “should” be, but the way it was forged.
Conceptually, RaiSR RuiNS sits in the lineage of glitch studies. The Dutch theorist and artist Rosa Menkman argued in her manifestos and writings that artifacts — compression errors, codec failures, resolution anomalies — are evidence about the media systems that produce them. De Wilde performs a comparable archaeology on the internals of machine learning: checkerboards are traces of convolutional stride; blurs are not painterly softness but statistical indeterminacy. By enlarging these traces beyond their intended scale, the project transforms transient errors into monumental surfaces. The works resemble architectural detritus, eroded frescoes, or geological strata — objects in which process has become the portrait.
To place RaiSR RuiNS historically: Generative Adversarial Networks were introduced in 2014 by Ian Goodfellow and colleagues. The architecture — a generator learning to fool a discriminator in adversarial play — quickly became a laboratory for image synthesis. The 2015 DCGAN paper by Alec Radford, Luke Metz and Soumith Chintala demonstrated that convolutional networks could produce coherent images from unlabeled data, but outputs remained small (often 64 × 64 pixels) and visibly computational: doubled eyes, melting limbs, repeated textures.
Between 2016 and 2018 GAN art moved from labs into broader cultural view. Methods such as PPGN (2016) added classifier guidance; NVIDIA’s Progressive Growing of GANs (2017–2018) pushed resolution and realism further; and public milestones — Robbie Barrat’s 2018 experiments and the 2018 Christie’s New York sale of a work by the collective Obvious — marked the medium’s market visibility. Yet in this formative period the seams of computation were legible; early GAN image making wore its machinery on its sleeve.
RaiSR RuiNS embraces that legibility and enlarges it until the machinery becomes landscape, and we can observe that the machine’s imagination left fingerprints.
BIO Frederik De Wilde is an artist whose practice operates at the intersection of art, science, technology, and design. His work investigates the inaudible, the intangible, and the invisible—probing both digital and physical realms to reveal how technological systems condition perception, reorganize environments, and quietly shape social and planetary realities. Through speculative research and sustained engagement with complex systems, De Wilde develops conceptual and aesthetic frameworks that confront contemporary ecological and technological urgencies.
He first gained international recognition for his pioneering Blackest-Black nano-engineered artworks (2010), developed in collaboration with Rice University and NASA. Conceived as an exploration of “nothingness,” the project received the Next Idea Award at Ars Electronica and the Best European Creative Cities Award, and was widely featured in international media, including Huffington Post, Dazed, and TED’s Ideas Worth Spreading. The work catalyzed renewed discourse around ultra-black materials and the aesthetics of absence within contemporary art.
Experimentation has remained central to his trajectory. Early projects include research into interspecies communication with weakly electric fish (2006), and the generative NRS series (2010), which the artist described as “painting with data” to render landscapes of the information age. In 2013, his collaboration with the Australian National University’s Department of Quantum Science led to some of the first artworks generated through true quantum randomness, translating subatomic indeterminacy into material form.
In 2016, De Wilde presented the world’s first AI- and evolutionary-algorithm-generated 3D-printed artworks, exhibited at Ars Electronica, alongside an earlier interactive neural-network installation (2015) now held in the permanent collection of the Frost Art Museum. His research has also yielded an AI-encoded Dazzle camouflage system that deploys evolutionary computing to disrupt metadata labeling and computer-vision pipelines, challenging regimes of algorithmic visibility.
De Wilde has published internationally, including the peer-reviewed article “Artistic Approaches to Design and Manufacturing Techniques Dedicated to Space Applications” in Leonardo (MIT Press, 2019). His distinctions include finalist nominations for the Giant Steps XPRIZE Lab at Massachusetts Institute of Technology, the ZKM Center for Art and Media Karlsruhe App Art Award, the TED Worldwide Talent Search (2013), and the Arab Bank NFT Prize. In 2024, he received first prize in the Herbert W. Franke Contact Attempt competition, and in 2025 he was selected for the Serpentine Future Art Ecosystems × Alias Studio Residency and the objkt labs Residency.
His works are held in the collections of the ZKM Center for Art and Media Karlsruhe, the Smithsonian Institution, and the Prince Albert II of Monaco Foundation, among others. De Wilde has exhibited internationally at major institutions including the Venice Biennale, BOZAR Centre for Fine Arts, MAAT – Museum of Art, Architecture and Technology, ArtScience Museum, Centre Pompidou, and the Carnegie Museum of Art. He has also been featured in Hyundai’s ART+TECHNOLOGY series produced with Bloomberg Media.