In June 2022 the comedian Guy Kelly typed a word he had just made up — crungus — into DALL·E mini, and the model answered nine times with the same snarling, hairy goat-demon. Crungus Amongus asks every text-to-image model it can reach the same question, and keeps the answers. It is a Cybernetic Studio project by Ben Swift.

Ten images of a hairy, horned goat-demon with bared teeth, generated by min-dalle from the single word crungus.
The original crungus, re-summoned in 2026 from min-dalle: the PyTorch port of the model behind DALL·E mini, still running on rented GPUs four years on. Ten samples from the one word, with nothing else in the prompt.

A word with no referent#

The first crungus was uncanny because it was repeatable: nine samples, one creature. A word with no meaning had a stable referent inside the model, and the internet named it the first digital cryptid. When Craiyon shipped its next version in 2023, the creature was gone.

A meaningless word makes a good probe because it gives the model nothing to work with. There is no concept to retrieve, so what comes back is whatever the model’s training left in the space around the word. Every model has a crungus in this sense, even if its crungus is a mushroom, or a brand, or a texture, and the archive exists to watch that change from one generation of models to the next.

The archive#

Crungus Amongus puts the question to every model in Replicate’s text-to-image collection, plus a hand-curated set of relics reaching back to 2021. Each model gets the same two prompts, “crungus” and “a picture of a crungus”, ten images apiece, with no style words, seeds or negative prompts. The prompt set is fixed, because changing it would break the comparison across years.

At the time of writing the archive holds 1,594 images from 83 models. Eight more were asked and produced nothing. Old models rot: weights vanish, hardware retires and safety filters refuse, and the ones that failed are listed on the site alongside the ones that answered.

Six images labelled 2021 to 2026: a psychedelic texture, a brown segmented thing with green teeth, a horned toad, a green ogre, an orange furry imp with horns, and a rotting hairy demon.
The most typical crungus of each year: the image with the highest mean similarity to every other image from that year's models. A CLIP-guided texture in 2021, then an unstable run of toads, ogres and imps, and in 2026 the goat-demon is back.

Does this model have a crungus?#

Beside each model the site reports a consistency score: the mean pairwise similarity of the ten images’ CLIP embeddings. High means the word lands somewhere specific inside the model; low means each sample is a fresh guess. It says nothing about image quality, and a model that always paints the same beige void would score perfectly. The number answers one question only: whether the model has a crungus.

Three rows of ten images. Top: ten identical images of a horned red creature in a forest. Middle: ten near-identical photographs of a toadstool. Bottom: ten unrelated images including a logo, a cartoon dinosaur, a bowl of soup and an American footballer, several with the word CRUNGUS written on them.
Three ways to score. Top: SDXL Lightning returns the same image ten times, and tops the table without meaning much by it. Middle: for Qwen-Image the crungus is a toadstool, every time. Bottom: HiDream has no crungus and treats the word as a brand, printing it on posters, packaging and a football jersey.

The loop#

The archive’s oddest finding is that the goat-demon came back. Between 2023 and 2025 the models mostly drew something else, and then in 2026 OpenAI’s image model produced the original creature again, snarling in a dim corridor, often with its name lettered across the top like a film poster.

Two rows of ten images. Top: the 2022 min-dalle goat-demon. Bottom: a 2026 model's version of the same creature, several with CRUNGUS written in the image.
Top: min-dalle, 2022. Bottom: gpt-image-1.5, 2026. The creature is recognisably the same, and the later model knows what it is called.

The archive cannot prove the route, but it is hard to see another one. In 2022 a model hallucinated a creature from nothing; people screenshotted it, wrote it up and gave it a Wikipedia article; and the write-ups went into the training data of the models that followed. That is a feedback loop, output fed back as input, and it is why the project keeps a history. Current model comparisons run the same prompts on the same day and keep none of it, and the loop only shows up over years.

The atlas#

Every image in the archive is embedded with CLIP and projected to two dimensions, so that images which look alike sit together. Clusters are creatures: where many models agree, a crungus exists. Only local distances mean anything. The clusters are real; the gaps between them are an artefact of the projection.

A dark field scattered with hundreds of small thumbnails, gathered into clusters: demons and ogres at upper left, cute creatures at upper right, animals and mushrooms in the middle, and a cluster of posters with the word CRUNGUS at the bottom.
All 1,594 images by visual similarity. The goat-demons and their descendants gather at the upper left, the toys and cute creatures at the upper right, the mushrooms and animals between them, and the models that read the word as a brand name form their own island at the bottom.

The archive is live at crungusamong.us, with a page for every model and a pan-and-zoom version of the atlas. The pipeline and site are open source at github.com/ANUcybernetics/crungus-amongus.

Other projects#

You are on Aboriginal land.

The Australian National University acknowledges, celebrates, and pays our respects to the Ngunnawal and Ngambri people of the Canberra region and to all First Nations Australians on whose traditional lands we meet and work as the oldest continuing culture and knowledges in human history.

arrow-right bars search times