How to prompt Qwen-Image: put the words in quotes, then tell it where they go

Qwen-Image is Alibaba's image family, and its headline strength is the thing most image models are worst at: rendering words correctly, multi-line and in two languages. On LUVI it spans ten models—Qwen Image 2.0 and 2.0 Pro, Plus, Max, and the Edit and Edit Plus tiers. On September 22, 2026 a single image quoted 42 credits on the Plus tier, 105 on Max and 120 on 2.0 Pro. This guide is about the one habit that makes the family worth choosing.

How do I get exact words into the image?

Put every visible string in double quotes, then say where it sits and how it looks. The quotes are what tell the model these are words to paint rather than a description of words; the position and the treatment are what stop it from putting them somewhere else.

A narrow corner coffee shop on a rainy evening, shot straight on at eye level. Above the doorway a weathered enamel sign reads "COLD BREW" in cream condensed lettering, centered in the top third. Below it, a small chalkboard on the wet pavement reads "OPEN TILL LATE" in white handwriting. Warm interior light spills onto the pavement, neon reflections in the puddles, commercial photography, shallow depth of field.

Keep the quoted words exactly as you want them—same language, same capitalization—and never add a string the picture does not need. The model paints whatever is quoted, so a stray word in quotes becomes a stray word on the wall. Fewer simultaneous strings also render cleaner: two well-placed zones beat five competing ones.

Around the text, the prompt follows a plain order: subject, setting, style, camera (shot size, angle, lens, composition), atmosphere, details. Spatial relationships are read literally, so "center-right", "top third" and "below it" do real work.

How do I change text that is already in an image?

You write it as a replacement, with both strings in quotes. Alibaba's own edit guide gives the pattern in its examples: "Replace 'HEALTH INSURANCE' on the Scrabble tiles with 'Tomorrow Will Be Better'." and "Change the phrase "Take a Breather" on the note to "Relax and Recharge"."

Type treatment is a separate instruction, and the guide's examples show how far it goes: changing a word to "a black ink-drip font", "a black handwriting font", "a black pixel font", to red, to a blue-purple gradient, to candy colors—or changing its material to metal. One change per edit reads more reliably than a list.

Do negative words work?

Not in the prompt body. "No watermark" and "without people" are not read there; they simply add the words watermark and people to what the model is thinking about. What exists instead, on some models, is a separate field.

This varies inside the family, so it is worth checking before you rely on it. On LUVI, Qwen Image Max carries a negative_prompt parameter—the place for "blur, extra fingers, garbled text"—while Qwen Image 2.0 Pro exposes only a size control. The help center keeps the general picture of when a negative prompt exists across the catalog.

How do I point at more than one uploaded image?

By array position. Alibaba's rule is literal: "Image numbers in prompts correspond to array position: the first image is 'image 1', the second is 'image 2'", and bracketed markers like "[image 1]" work too. The identifier then acts as a noun in the sentence, which is what makes multi-image edits readable: "The girl in Image 1 wears the necklace from Image 2 and carries the bag from Image 3 on her left shoulder."

Write the operation, not a description of the finished picture—"Change the hair to a short brown bob. Keep the face, blouse and lighting as in the original."—and say explicitly what must stay. A single input image cannot exceed 10 MB.

What about size, count and language?

Three practical details, and the first one catches people.

• size is written WIDTH*HEIGHT with an asterisk, not an "x". Qwen Image Max takes five values, from 1664*928 through 1328*1328 to 928*1664; Qwen Image 2.0 Pro takes a free size string and defaults to 1024*1024. A value in the wrong format is dropped and the default is used instead.
• num_images on Max generates 1 to 4 images in one request.
• Language: Simplified Chinese and English. Alibaba's FAQ is direct about it—"The model currently supports Simplified Chinese and English. You can try other languages, but performance is not guaranteed." A string in another language that must render should still be quoted and left untranslated, but expect less reliability.

Keep the whole prompt dense rather than long; under about 200 words is the working shape for a single image.

How we checked

• Models: Qwen Image Max and Qwen Image 2.0 Pro Text-to-Image, with the Plus, 2.0, Edit and Edit Plus tiers read from the catalog.
• Settings: one image at 1664*928 on Max and 1024*1024 on 2.0 Pro.
• Date and what was read: September 22, 2026—the models' parameter contracts and live credit quotes in LUVI, LUVI's own prompting guide for the Qwen-Image family, and Alibaba Cloud's Qwen-Image-Edit guide (last updated September 11, 2026).
• Results: 105 credits on Max and 120 on 2.0 Pro; 42 on the Plus tier. Max exposed a negative_prompt field and 2.0 Pro did not. No outputs were generated for this post, so it makes no claim about output quality.
• Drawbacks: the advice to keep the number of simultaneous strings low comes from LUVI's prompting guide rather than a published measurement.

In LUVI

Pick a model from the Qwen-Image family page, write your strings in quotes with a position each, and run the feather button beside the prompt box before you generate—it applies LUVI's prompting guide for the model you selected, including the rule about negative words in the body. If you are sending the file somewhere with a fixed shape, set size deliberately rather than accepting the default, and remember the asterisk.

New accounts start with 200 credits, and verifying a phone and a card adds up to 1,500 more. You can create a LUVI account and find out how your own copy renders.

Sources

• Qwen-Image-Edit guide — Alibaba Cloud Model Studio, official documentation, last updated September 11, 2026. Accessed September 22, 2026.