Getting Started with Qwen Image: Chinese vs. English Prompts and In-Image Text Rendering

Qwen (Tongyi Qianwen Image) is an image generation model from Alibaba Cloud. Its training covered massive amounts of both Chinese and English text, so it usually understands natural Chinese descriptions well. You do not need to translate your idea into English first; describe the scene directly in Chinese and the model will generally grasp the main intent.

For Chinese users this clearly lowers the barrier: you can describe an image to Qwen the same way you would describe a photo to a friend, without first learning a special set of English keywords. Rather than forcing your idea into English and worrying about distortion in translation, writing in your native language usually stays closer to the picture you had in mind.

Early models had limited natural language understanding, so the English-speaking community developed a keyword-stacking style: long comma-separated lists of short phrases and quality boosters, with words like masterpiece, best quality and ultra detailed repeated over and over. That habit existed because those models could only trigger concepts through keywords; with Qwen it is not required.

Newer models are generally much better at understanding full sentences. When you write in Chinese, simply organize subject, action, environment, lighting and composition into coherent description. There is no need to pile up modifiers, and long strings of quality boosters usually bring no obvious benefit.

Two general tips: order your description as subject, action, environment, lighting, then style, since content placed earlier usually carries more weight; and state the details you care about explicitly, which works better than repeating synonyms.

For subjects rooted in Chinese culture, Chinese descriptions are usually more reliable: clothing terms like hanfu, mamian skirt and pankou buttons; architecture like upturned eaves, classical gardens and siheyuan courtyards; foods like candied hawthorn and xiaolongbao; and painting concepts such as ink wash and blue-green landscape. A single Chinese word lands exactly, while an English translation can easily drift.

When the image must contain Chinese text, such as shop signs, poster headlines, subtitles or calligraphy, writing in Chinese is essential: if you describe a poster with Chinese characters in English, the model usually has no idea which characters to write. Only by putting the actual copy into the prompt in Chinese does the model get a chance to render it correctly.

Conversely, for Western subjects or style vocabulary popular in English-speaking communities, English prompts are perfectly reasonable. You do not have to pick one language forever; switch flexibly depending on the subject.

The core trick is to wrap the copy in quotation marks and state explicitly that something says or displays it, for example: a wooden shop sign that reads 「老字号茶馆」 (an old-brand teahouse). Text inside quotes is generally treated as copy to be rendered exactly.

The same applies in English: wrap the copy in quotes, as in A wooden sign that says "OPEN". Chinese copy inside quotes can render too, but keeping the whole prompt in one consistent language usually gives a higher success rate.

Keep the amount of text small: fewer characters are easier to get right. Signs of two to six characters and poster headlines of no more than ten generally perform better; even when long paragraphs are spelled correctly, the layout often breaks. You can also specify where the text sits (top, centered, lower left) and the font feel (brush handwriting, song typeface, rounded, neon glow); the model usually makes use of these details.

If the text comes out wrong, do not start from scratch: move the copy earlier in the prompt, add a line stating the text must be accurate, and generate again. Generating several images and picking the best one is normal practice.

Lighting can be described in everyday Chinese: backlit at dusk, soft morning light, harsh noon light, window side light, candlelight, neon light, stage spotlight, plus rim light (highlights outlining the subject) and volumetric light (beams passing through mist or dust). The model generally understands all of these phrasings.

Camera and composition have direct Chinese equivalents too: close-up, medium shot, half body, full body, extreme long shot, low angle looking up, overhead shot, first person view, shallow depth of field with a blurred background, and macro. These are everyday photography words; there is no need to swap them for English terms.

For a cinematic look, try cinematic color grading, widescreen aspect ratio and shallow depth of field; for a realistic look, try natural light, film grain, or the feel of a casual phone snapshot.

Portrait example: an Asian woman in her twenties, wearing a cream white knit sweater, sitting by a window filled with afternoon sunlight, soft window side light falling on her face, close-up shot, shallow depth of field with a blurred background, film-like texture, warm tones.

Landscape example: an early morning in a Jiangnan water town, whitewashed houses with dark tiled roofs lining both banks of the canal, a stone arch bridge crossing the water, thin mist drifting, a small boat gliding slowly by, soft morning light, extreme long shot composition, tones blending blue-green landscape painting with modern photography.

Text poster example: a vertical guofeng tea poster, the subject a steaming cup of jasmine milk tea, the background pale cyan ink-wash mountains, four brush-written characters 「山间茶事」 (Tea in the Mountains) centered at the top, a smaller line 「秋日上新」 (New for Autumn) below, an elegant palette with generous white space.

When the text comes out wrong, shorten the copy first, move it closer to the front of the prompt, spell out the font and position, then generate several images and pick one. Text rendering generally cannot succeed every single time yet; that is normal, so do not assume your prompting is at fault.

When the image does not match your expectation, check whether the subject is drowned by too many modifiers and move the most important content to the front of the sentence; change one thing at a time and regenerate, so you can tell which word caused the change.

Start by imitating: pick reference images you like on the site, study the structure of similar Chinese prompts, then swap in your own subject and scene. After a few rounds of practice you will develop a prompting habit that suits you.