Should Prompts Be Written in Chinese or English?
Fixing a bug at 2 a.m., halfway through, this question comes up again — one I get asked a lot: should prompts be written in Chinese or English? The answer is the same as asking "which programming language is better" — it depends on the environment you're running it in, not on which language is inherently more advanced.
It depends on the model — don't apply one rule to everything
Nano Banana, Seedream, and similar models.
Both of these have solid Chinese semantic understanding, especially Seedream, whose training corpus leans heavily Chinese. Write "a hutong alley at dusk, a cat perched on the wall," and it correctly understands culturally loaded words like "hutong" and "wall perch" — translating that into English risks losing detail, because English simply doesn't have a one-to-one equivalent for some of these terms. For this kind of model, write Chinese directly, no need to go around it.
Midjourney.
This one's the opposite. MJ's training data and parameter ecosystem have accumulated the most community experience around English prompts — style words, camera terms, lighting terms are almost entirely built on an English jargon system (things like cinematic lighting, bokeh). Force-translate that into Chinese and the model gets the gist, but a lot of these conventionalized style descriptors either have no Chinese equivalent or lose effectiveness in translation. For this kind of platform, use English wherever you can, especially for the style and parameter portions.
Flux.
Similar logic to MJ — trained mostly on English-language data, so English prompts respond more reliably.
GPT-based image generation and similar models.
These take the natural-language-understanding route, and the gap between Chinese and English isn't as wide, because they're language models at heart, with more balanced comprehension across both. For these, write in whichever language feels natural to you — no need to agonize over it.
A pitfall I fell into myself
I once forced a very literary Chinese description into an English translation and fed it to a Chinese-friendly model — the result came out completely off. That's because the act of translating itself is a loss of information, especially for those evocative, mood-laden adjectives — translating Chinese into English basically means rewriting the meaning from scratch. It clicked for me afterward: the problem wasn't picking the wrong prompt language — it was the unnecessary extra translation step, adding a layer of noise for nothing.
Roughly speaking: if you can write directly in the language the target model is actually good at, don't think it through in one language first and then translate — that translation step almost never pays off.
A mixing trick that covers both bases
If you're not sure which language a model prefers, there's a middle-ground move: write the subject and scene description in whichever language feels most natural to you (usually Chinese, since you'll nail the details more accurately), and use English for style and technical vocabulary (terms like cinematic, bokeh, film grain — these are internationally standard photography terms that tend to lose meaning in translation anyway).
For a portrait description, for example: "a woman in a Xiangyunsha qipao, sitting in a rattan chair, cinematic lighting, shallow depth of field" — the subject detail is accurate in Chinese, the style words are idiomatic in English, and neither side loses out.
E-commerce scenarios are actually easier to judge
For e-commerce product descriptions, regardless of Chinese or English, models handle structured instructions like "pure white background," "product centered," "no shadows" pretty consistently. Language choice has little impact in this scenario to begin with, so don't overthink it — just use whichever language you type faster in.
One sentence isn't enough to wrap this up, so here are two
Picking a language really comes down to picking "whichever path is closer to the model's training corpus." Domestic models lean closer to Chinese, international ones lean closer to English, natural-language models are close to both. If you genuinely can't tell, go to the prompt tool and try both — it takes five minutes, which is worth a lot more than half an hour of agonizing over it. Same as debugging — don't guess, just run it and check the log.