Won Ton A Cat Tale Told In Haiku

The Lost History of Won Ton and Why People Keep Asking About It

Won Ton: A Cat Tale Told In Haiku was a real piece of art-science software from around 2011, created by researcher David McCallum at Parsons The New School for Design. It used neural networks to generate abstract images and described each output in a short, haiku-like statement about a virtual cat. It went viral on Reddit and was featured in the Whitney Biennial in 2012 before disappearing into the archives of forgotten AI art tools. At its core, Won Ton was a generative adversarial setup before GANs became mainstream. It used a self-organizing map — a type of neural network that clusters similar data together — trained primarily on cat images from the internet. The system would then generate novel images by exploring the latent space between those clusters, and each generated image was accompanied by a haiku written from the perspective of a fictional cat character named Won Ton. The whole output looked something like this: an abstract, often unsettling image rendered in warm tones, followed by two or three lines of text describing what the "cat" was experiencing. The haikus ranged from mundane ("sleeping in the sun patch") to bizarre ("the box has teeth today"). That juxtaposition between the visual abstraction and the deadpan feline narration was the entire artistic concept.

Why It Matters in the Context of Generative Art

Most people who stumble onto Won Ton now are doing so because they read about early AI art history or encountered a reference in a contemporary generative art project. The system is notable because it predated the current wave of neural style transfer, DeepDream, and DALL-E by roughly five to seven years. McCallum's work demonstrated that you could make a generative system both visually interesting and conceptually framed through narrative — an approach that modern AI art tools largely abandoned in favor of pure prompt-based control. The technical pipeline was straightforward for its time. You fed it thousands of cat photographs scraped from the web. A Kohonen self-organizing map clustered them by visual similarity. Then the system would wander through those clusters, interpolating between them to create images that didn't exist in the training set. The haiku descriptions were generated separately, often by McCallum himself or by a simple text template system, which is why the quality varied so much between outputs.

How to Experience It Today

There is no official working download available. The original project page at won-ton.org was taken down around 2014, and the source code was never formally published in a runnable state. What exists are: Archive.org snapshots of the original website, which contain the gallery of generated images and their accompanying haikus. You can view these at the Wayback Machine by searching for won-ton.org. A few academic papers that describe the architecture in enough detail that someone with neural networking experience could attempt a reconstruction. McCallum presented the work at several conferences including Leonardo Electronics Auction and various digital art symposia.

Imitation projects on GitHub. Several developers have built simplified recreations using Python and libraries like TensorFlow or PyTorch, typically using a self-organizing map on a dataset of their choice rather than specifically cat images. I tried to reconstruct it myself around 2018 using a SOM implementation and a custom dataset. The main problem I ran into was that the original system seemed to have used a specific preprocessing pipeline for the cat images that nobody documented — things like resolution normalization, color space conversion, and likely some kind of feature extraction layer before the clustering step. Without that, the generated images just looked like noisy gradients with no coherent structure. The workaround was to use a pre-trained convolutional neural network as a feature extractor instead of raw pixel data, which produced results much closer to the aesthetic of the original Won Ton outputs, though still not identical.

The Common Misunderstandings

People often assume Won Ton was an image-to-image translation tool or a style transfer program. It wasn't. It was purely generative — it created new images from learned patterns rather than transforming existing ones. Another frequent confusion is that the haikus were AI-generated. They weren't, at least not in the original deployment. The text component was hand-written or template-driven, which is an important distinction because it means the "personality" of Won Ton came entirely from the artist's curatorial choices, not from a language model. A more practical issue: the original visual output had a very specific warm, slightly sepia-toned palette that many recreations fail to reproduce. This wasn't an accident of the neural network — it was a deliberate post-processing step that applied a color grading overlay. If you're building a recreation and the colors look too cold or too neutral, adding a warm tone curve in post will bring it much closer to the original aesthetic.

What Won Ton Predicted That Modern AI Art Still Struggles With

The most interesting thing about Won Ton isn't the technology — it's the concept. McCallum understood early on that raw generative output needed narrative framing to feel meaningful to viewers. Modern tools like Midjourney and Stable Diffusion have mostly solved the image quality problem but have barely touched the narrative problem. A generated image is still just an image unless someone adds context. Won Ton wrapped its outputs in haikus specifically to force that connection between the abstract and the personal. The system also anticipated the current conversation about AI art and authorship. By giving the generator a fictional cat persona, McCallum was creating a humorous but pointed commentary on how we anthropomorphize AI systems. We're still doing that today with chatbots and image generators, just without the haikus. If you want to explore the original work, the archive.org snapshots are the most reliable source. The gallery alone contains enough material to understand the scope of what McCallum was attempting. For anyone interested in the technical side, the conference presentations are more useful than the academic papers, since they include actual demonstrations of the system in operation rather than just architectural diagrams.