One Scene, Six Languages: What Gets Lost When You Prompt in Your Own Words

Describe a rainy evening by a window and the words you reach for depend on the language you think in. That is not a flaw; it is why your own words feel right. But image models tend to respond best to short English descriptors, so the gap between how you think and what the model reads is where detail gets lost.

Here is one scene described six ways, and what each language teaches us about writing better prompts.

The scene

Six ways to say it

What each language does differently

• Korean: modifiers come before the noun and the verb comes last, so a description is a chain: 창가에서 책을 읽는 여자 is literally "woman who reads a book by the window". And one adjective like 은은한 covers soft, subtle and faint at once. An English tag has to pick one.
• Japanese: mimetic words carry texture and mood. しとしと is rain falling quietly and steadily, ふんわり is soft and airy. English has no short tag for either, so you describe the visual result instead: light drizzle, soft airy glow.
• Chinese: compact and image-driven. 细雨 (fine rain) takes two characters, and set phrases like 烟雨朦胧 (hazy, misty rain) pack a whole mood into four. That is lovely to write and hard to turn into one tag.
• Korean, Japanese and Chinese together: none of them requires articles or plural markers, so "one woman or several?" often stays implicit. In English, say it out loud: 1woman, solo.
• Spanish: adjectives usually follow the noun and agree with it (luz suave, luces suaves), and the verb already says who is acting, so the pronoun disappears. English-style stacking such as soft diffused rim light has no word-for-word equivalent.
• Brazilian Portuguese: structurally close to Spanish, but with its own vocabulary. Garoa is a fine, steady drizzle, and entardecer is dusk as a process (the afternoon turning to evening) rather than a clock time.
• English: noun phrases stack naturally (soft backlight), which is why tag lists work so well. But English has near-synonyms with different looks too: dusk, twilight and golden hour are not interchangeable.

One word, several tokens: the real translation problem

A translator hands you the most common English word. But for a prompt, the interesting question is which of several candidates matches your picture:

Choosing between those candidates is the creative decision. Don't leave it to a translation engine; make it yourself, once, and save the answer.

A method that works in any language

1. 1Write the scene in your own language first, as short phrases rather than one sentence.
2. 2For each phrase, list two or three English candidates. There is rarely a single right one.
3. 3Compare against a real example image and pick the candidate that looks like what you meant.
4. 4Save the winners as a personal glossary: your phrase, the English token, and a note on why.
5. 5Next time, reuse the glossary instead of translating from scratch.

The reasoning behind working one concept at a time is in Stop Using Google Translate for AI Prompts.

How mysnipbox fits in

When mysnipbox decomposes an image, it produces short English tags. The Multilingual Token Bridge shows each tag next to its translation in the language you pick, and caches the pairs in your local database. A quick list covers Korean, English, Japanese, Chinese, Spanish and Portuguese, and a searchable list covers many more. Tag Insights then shows what a tag looks like in your own images, which is step three of the method above, inside your library.

Translation runs through your own cloud key or a small local text model, so your tags don't have to leave your machine. For Korean, Chinese and Japanese, qwen2.5:3b is a good light choice; for Spanish and Portuguese, llama3.2:3b. See Local models (Ollama) for sizes.