The Animal Face Dictionary: Every Term You'll Meet, Explained

Animal face culture comes with a vocabulary — part Korean beauty slang, part psychology, part machine learning — and half the confusion newcomers feel is really just missing definitions. This dictionary collects every term you are likely to meet on this site or in face-type discussions online, organized by theme. Bookmark it; every other article here will make more sense with it.

The Korean Core Vocabulary

• -sang (상 / 相). Suffix meaning "face" or "physiognomy." Attached to any noun it means "the [noun] type of face": gangaji-sang, goyangi-sang. If you learn one particle of Korean beauty slang, learn this one.
• Gangaji-sang (강아지상). Puppy face — the warm, round-eyed, approachable archetype; the founding category of the whole trend. See the dog face guide.
• Goyangi-sang (고양이상). Cat face — the composed, almond-eyed, elegant counter-archetype. See the cat face guide.
• Yeou-sang (여우상). Fox face — sharp, witty, strategically charming. The term carries centuries of fox-spirit folklore behind it.
• Gongnyong-sang (공룡상). Dinosaur face — strong-boned, striking, charismatic. Coined as fan slang and worn today as a badge of honor.
• Aegyo-sal (애교살). Literally "charm fat": the small pouch of muscle directly under the eye that plumps when smiling. The single most discussed feature in Korean beauty content — prominent aegyo-sal is the signature of the warm cluster, and an entire makeup category exists to fake it.
• Dongan (동안). "Baby face" — a face that reads younger than its age. Covered in depth in our dongan article.
• No-an (노안). The opposite of dongan — a face that reads older than its age. Less insulting than it sounds; often applied to mature-featured teenagers who grow into celebrated dinosaur or wolf types.
• Insang (인상). "Impression" — the overall feeling a face gives, as distinct from its individual features. The animal face framework is fundamentally a taxonomy of insang.
• Gwansang (관상). Traditional Korean physiognomy — the centuries-old practice of reading fortune and character from facial features. The animal face trend borrows its vocabulary but not its fortune-telling claims; the full story is in our gwansang introduction.
• Gumiho (구미호). The nine-tailed fox of Korean folklore — the mythological ancestor of the fox face's dangerous-charm reputation.

The Ten Archetypes at a Glance

• Dog — warm, loyal, universally approachable. The trust archetype.
• Cat — elegant, composed, magnetic at a distance. The mystique archetype.
• Fox — sharp, quick, playfully strategic. The wit archetype.
• Rabbit — bright, fresh, youthfully charming. The energy archetype.
• Bear — steady, cozy, quietly strong. The stability archetype.
• Deer — graceful, gentle, quietly captivating. The elegance archetype and the framework's bridge type.
• Hamster — round-cheeked, cheerful, instantly endearing. The comfort archetype.
• Wolf — intense, selective, quietly confident. The focus archetype.
• Dinosaur — strong-featured, memorable, cool outside and warm inside. The presence archetype.
• Chick — light, optimistic, impossible to dislike. The freshness archetype.

Framework Concepts

• Warm cluster. Dog, rabbit, hamster, chick, bear — the family of soft-featured, approachable types. Warm faces earn trust fast and authority slowly.
• Sharp cluster. Cat, fox, wolf, dinosaur — the family of defined-featured, composed types. Sharp faces earn authority fast and trust slowly.
• Bridge type. The deer face's special status between the clusters — gentle enough for the warm family, refined enough for the sharp one.
• Hybrid type. A result with two strong scores, read as a blend (cat-fox, dog-rabbit, wolf-dinosaur). The majority of real results are hybrids; fan culture treats famous combinations as types in their own right.
• Cross-cluster blend. A hybrid whose two types come from opposite clusters (dog-wolf, rabbit-cat) — the rarest pattern, usually meaning bone structure and expression are sending different signals.
• Core / social / structural impression. The three reads produced by testing multiple photos: the type that always appears, the type that appears when smiling, and the type that appears when neutral. Explained in Why Results Change.
• Puppy–cat axis. The original two-camp debate that the ten-type framework grew out of; most types can be located on it.

The Psychology Terms

• First impression. The trait judgments (trustworthy? competent? warm?) observers form from a face — measurably within a tenth of a second of exposure in laboratory studies.
• Thin-slicing. Psychology's term for judgments made from tiny samples of information, faces included. Fast and confident, but far less accurate than the confidence suggests.
• Halo effect. The bias where one positive impression (an attractive or warm face) inflates unrelated judgments (honesty, intelligence). The engine behind most face-based assumptions.
• Baby schema (Kindchenschema). The set of infant features — large eyes, round cheeks, small chin — that automatically triggers warmth in observers. The scientific cousin of the warm cluster.
• Physiognomy. The historical claim that character can be read from facial features. Repeatedly tested, repeatedly failed; the reason this site labels itself entertainment. See our ethics article for the full discussion.

The Technology Terms

• TensorFlow.js. Google's open-source machine learning library for JavaScript — the engine that runs this site's model entirely inside your browser.
• On-device / in-browser AI. The architecture where the model computes on your hardware and your data never reaches a server. The reason your photo is never uploaded here.
• Image classification. The task the model performs: assigning an input image scores across a fixed set of categories — here, the ten archetypes.
• Confidence score. The percentage next to each type: the model's estimate of how well your photo matches that category's learned pattern. All ten always sum to 100%.
• MobileNet. A lightweight neural network architecture designed for phones and browsers; the visual backbone our classifier is built on.
• Transfer learning. Training technique where a network that already understands general images is adapted to a specific task with a curated dataset — how a browser-sized model can classify faces at all.
• Teachable Machine. Google's tool for training small classification models from labeled examples; the platform our model was trained with.
• Inference. The moment of actually running the model on your photo — on this site, a few hundred milliseconds inside your own browser tab.

Using the Vocabulary

A worked example, fully translated: "She's a classic goyangi-sang but with strong aegyo-sal, so her result is probably a cat-dog cross-cluster hybrid — cat structural impression, dog social impression." Translation: her bone structure reads elegant and composed, but her under-eye fullness warms her smile so much that smiling photos will score toward the warm cluster. If that sentence now parses, you are fluent.