Your Negative Prompt Can’t Fix the Wrong Culture. Only Positive Detail Can.

TL;DR — Trying to generate a Korean historical figure, the output kept coming back Chinese. Putting chinese, hanfu in the negative prompt did nothing. What fixed it was naming four small positive features the costume actually has — a white collar band, two chest ties, a straight collar line, a waist-seamed pleated skirt. A generative model fills underspecified regions with its own prior, and you cannot subtract your way to a detail that was never described. The same failure, in a different costume, produced a rubber dinosaur and a duplicated subject.

This is the companion to a finding on the same rig about why colour words in a negative prompt cause oversaturation. Different mechanism, same lesson: the negative prompt is not where you fix things.

The failure

The brief was a Korean historical figure — traditional dress, a period courtyard. Round after round came back looking like Chinese wuxia. Not subtly: the robe, the collar, the roof eaves, the whole register.

The instinctive fix is to write down what you do not want. chinese, hanfu, wuxia in the negative prompt. It did not work, and it did not work repeatedly.

The reason is structural. A diffusion model does not hold a checklist of things to avoid; it holds a distribution over images that fit your description. "East Asian historical robe, dark, formal, courtyard" describes a region of that distribution containing far more Chinese imagery than Korean, because that is what the training data contains. The negative prompt shifts you away from a point, but if the description does not pick out Korean dress, the nearest thing you land on is still whatever dominates the neighbourhood.

A negation removes an option. It cannot supply a missing specification.

What actually worked: naming the discriminating features

Korean traditional dress is separated from its neighbours by a small number of visually specific parts. Naming them, in the positive prompt, as shapes rather than as terms, fixed it:

Feature	How to describe it	What goes wrong without it
Dongjeong (동정) — the collar band	A narrow white band wrapping the collar edge	The single strongest signal. One revision dropped the word white and the whole image reverted
Goreum (고름) — the ties	Two long ribbon ties at the right side of the chest	Fastening reads as a generic wrap robe
Collar line	Descends straight to the right armpit, not a deep crossed V	The deep X collar is the neighbouring silhouette
Cheollik (철릭) — the robe	Waist seam with a pleated skirt below it	Reads as a one-piece gown
Dancheong (단청) — building paintwork	Multi-colour geometric paintwork under the eaves	Silhouetted halls get upswept eaves in the wrong style

The hat deserves its own note, because it is the clearest case of the general problem. Asking for a gat by name produced a broad sedge or straw hat — a different garment from a different country. The name is in the model's vocabulary but not bound to the right shape.

Describing the geometry worked:

a flat circular disc brim with a narrow, tall cylinder rising from its centre, woven from horsehair so it is semi-transparent

That is the same object, specified as form. When a name fails, describe the shape. The model has certainly seen the object; it has not reliably learned your word for it.

The control that proves it is not "people"

Worth stating because the obvious conclusion from a run of failures is that the model is bad at this subject.

A batch of modern-dress Korean figures generated in the same session came out photographic, with no drift at all across four images. Same model, same pipeline, same nationality of subject.

So the failure is not people, and it is not Korea. It is period costume specifically — a category where the visual differences between neighbouring traditions are small, the training data is thin and mislabelled, and the model's prior is therefore both strong and wrong. Contemporary clothing is over-represented and unambiguous, so underspecification costs nothing.

That reframes when to worry. Ask for something the model has seen a million clearly-labelled times and vague prompts are fine. Ask for something where a handful of small features separate one tradition from another and every unspecified detail is decided against you.

The same failure wearing a different costume

A separate subject — a photorealistic dinosaur — failed three times in ways that look unrelated and are not.

Round one: a description fragment overrode the style instruction. The anatomy paragraph contained muted olive-brown hide. The look instruction asked for saturated colour. The concrete noun-phrase won and every image came out beige. Specific descriptive language beats general stylistic language, so a colour buried in an anatomy sentence quietly becomes the colour policy.

Round two: a style phrase requested the wrong render mode. With saturation raised, the output turned into what can only be described as a mobile-game 3D asset — clean, plasticky, floating. The cause was two clauses in a shared style block: digital art, vivid stylised illustration. Read literally that is a request for illustration, and the model complied. The same phrase was independently responsible for a flat, illustrative look in an entirely different project on the same rig.

Round three: specifying the reference class fixed it. Replacing the style with natural history documentary photograph and filling the skin description with physical detail — cracked keratin, neck folds, healed scars, dried mud, mucus at the nostril, saliva between the teeth, flies — removed the rubber texture. Not because the model learned anything new, but because "documentary photograph" selects a different region of the distribution than "digital art", and physical imperfections are what separate a render from a photograph.

And a bonus failure with a clean lesson. Describing the prey animal only by family — "a hadrosaur" — made the model duplicate the main subject instead of drawing a second species. A vague noun in a scene with one strongly-specified subject gets resolved as another copy of that subject. Two fixes, both positive specifications: name the actual genus, and state exactly one single <subject>.

The pattern

Four different-looking failures, one mechanism:

Symptom	What was underspecified	Fix
Wrong country's costume	The features that discriminate it	Name them positively, as shapes
Named object came out wrong	The word-to-shape binding	Describe the geometry instead of the name
Everything beige	Which colour statement wins	Remove the colour from the anatomy text
Plastic game-asset look	The render mode	State the reference class: documentary photograph
Second character is a clone	The second subject	Name the species; state the count explicitly

In every case the model was not refusing an instruction. It was resolving an ambiguity, and the prior it resolved with was the wrong one.

Underspecification is not neutral. Every detail you leave out is a detail the model decides — using the most common answer in its training data, which is exactly the answer you are trying to avoid when your subject is not the common case.

A practical procedure

Write the discriminating features, not the category. "Korean traditional dress" names a category the model has blurry boundaries for. A white collar band, two chest ties, a straight collar to the right armpit, a pleated skirt below a waist seam — those are the boundaries.

When a name fails, describe the shape. Names bind unreliably across languages and cultures. Geometry does not.

Check whether one sentence is quietly setting policy. A colour adjective inside an anatomy description, or illustration inside a style block, will beat a general instruction elsewhere in the prompt. Grep your own prompt for concrete nouns and adjectives you did not mean as instructions.

Say the render mode explicitly. Photograph, documentary still, film frame, engine render, illustration. If you do not choose, the style vocabulary you happened to use will choose.

Count your subjects. Any subject described more vaguely than the main one is a candidate to be redrawn as a copy of the main one.

Reserve the negative for defects. Blur, artefacts, watermarks, text — failure modes, not content and not culture. Everything else is a positive specification problem.

What this does not show

One model family, and the observation that negation underperforms positive specification is qualitative here — the negative prompt was tried and failed across several rounds, but no controlled A/B was run isolating it, and no measurement was taken. Treat "the negative did not work" as a reliable report of the attempts made, not as a measured effect size.

The list of discriminating features is what proved sufficient for these images. It is not a complete account of the costume, and a different scene or period would need different ones.

Both projects were abandoned for unrelated reasons before a final version, so nothing here is validated against a finished piece. The modern-dress control was four images.

FAQ

Why doesn't putting "chinese" in the negative prompt stop Chinese-looking results?

Because it removes a direction rather than supplying a description. If your positive prompt does not contain the features that identify the tradition you want, the model still has to fill those regions from its prior, and the prior is whatever dominates the training data for that description. Negation cannot add a detail that was never specified.

How do I get Korean traditional dress instead of Chinese?

Name the parts as shapes in the positive prompt: a narrow white band wrapping the collar, two long ties at the right chest, a collar descending straight to the right armpit rather than crossing in a deep V, and a waist seam with a pleated skirt below. For architecture, specify the multi-colour geometric paintwork under the eaves.

The model ignores the name of an object I want. What now?

Describe its geometry. A traditional Korean gat came out as a broad sedge hat when named, and correctly when described as a flat disc brim with a narrow tall cylinder rising from its centre, semi-transparent because it is woven from horsehair.

Why did my image come out looking like a mobile game asset?

Check your style block for digital art or stylised illustration. Those are literal requests for illustration. Naming the reference class you actually want — natural history documentary photograph, film frame — moves the output to a different part of the distribution, and adding physical imperfections (grain, scars, mud, moisture) is what separates a photograph from a render.

Why did a second character come out identical to the first?

It was described more vaguely than the first. A generic noun in a scene containing one strongly-specified subject tends to resolve as another copy of that subject. Name it specifically and state the count of the main subject explicitly.

---

← Back to all posts