Chrome ExtensionQwen Image Prompt Generator
Write prompts for Alibaba's Qwen-Image family, the models that render Chinese text in images better than anything else open. This generator follows the rules in Qwen's own prompt-rewriter source code rather than blog folklore: every literal glyph in quotes, flowing prose instead of tag lists, exclusions kept out of the positive prompt, and a warning built in about the language trip-wire that silently translates prompts.
Be specific: describe who/what is in the image with details
Quick Palettes
Generated Prompt
Fill in the form and click "Generate" to create an optimized Qwen-Image prompt.
Tip: Be as specific as possible with your subject description. Instead of "a dog," try "a golden retriever puppy with a red bandana."
Qwen-Image Tips
- • Put every literal glyph in double quotes. This is the single strongest rule and it comes straight from Qwen's own rewriter code
- • Write flowing prose, not comma-separated tags. The model was captioned on natural descriptive sentences
- • One Chinese character anywhere routes your whole prompt through the Chinese rewriter, so keep the body in one language
- • Never put negations in the positive prompt. Say what should be there and move exclusions to the negative prompt
- • For text, give position, exact wording, font, colour, size and medium in that order
- • Aim for around 200 words. The API silently truncates past its token cap rather than warning you
- • When editing, restate what must stay the same or the model will drift the face and clothing
The three rules that matter most
All three come from Qwen's published rewriter code and training-caption spec, not from community guesswork.
Quote every literal glyph
The rewriter instructs that every piece of displayed text be enclosed in double quotation marks, and quoted text is explicitly protected from translation. This is the difference between getting your words and getting an approximation of them.
A chalkboard sign reading "Qwen Coffee 2 dollars per cup", with a neon light beside it displaying "通义千问".
Write prose, never tags
Qwen-Image was captioned on natural descriptive text with no structured formats, and the rewriter forbids markdown lists, numbered items and headings. Tag soup is out of distribution for this model.
A coffee shop entrance at dusk, warm light spilling onto wet pavement, a hand-lettered sign propped by the door.
Keep negations out of the positive prompt
The Chinese rewriter explicitly bans negation words in the rewritten prompt. Describe what should be present and move everything you do not want into the negative prompt field, which Qwen does support.
Positive: a clean wooden table with a single ceramic bowl. Negative: chopsticks, clutter, extra utensils.
Watch out for the language trip-wire
Qwen decides which rewriter to run by scanning your prompt for any character in the CJK Unified Ideographs range. Finding even one routes the whole prompt through the Chinese rewriter, which rewrites it into Chinese and can Sinicize an English request entirely.
The fix is the pattern Qwen demonstrates itself: write the descriptive body in a single language, and put the exact glyphs you want rendered inside quotes. Quoted strings are protected and passed through untranslated, so an English prompt can safely ask for a sign reading "通义千问" without the rest of the prompt flipping language.
Qwen Image specs and configs
| Open weights | Qwen-Image and Qwen-Image-2512 for generation, Qwen-Image-Edit-2511 for editing. All 20B and Apache 2.0 |
|---|---|
| API only | Qwen-Image-2.0 from February 2026 and Qwen-Image-3.0-Pro from July 21, 2026. No weights, and 3.0 has no published parameter count |
| Architecture | 20B MMDiT with a Qwen2.5-VL text encoder for v1. The 2.0 report does not state parameter counts |
| Resolution, open weights | Fixed set from 1328x1328 square up to 1584x1056 at 3:2 |
| Resolution, API | Any width and height totalling between 512x512 and 2048x2048, roughly 4.19MP. Default 2048x2048 |
| Editing input | 1 to 3 images, 384 to 2048px per side, under 10MB each |
| Prompt limit | 800 tokens for most models, 1,300 for the 2.0 series, 4.5k for 3.0-Pro. Excess is truncated silently |
| Negative prompt | Supported, up to 500 characters. Negation words are not allowed in the positive prompt |
| Open-weight settings | 50 steps at true_cfg_scale 4.0 for generation, 40 steps for Edit-2511 |
| Other parameters | seed, n from 1 to 6, watermark toggle, and prompt_extend which defaults to on |
| Text benchmark | LongText-Bench 0.943 English and 0.946 Chinese for v1, independently replicated within a couple of points |
| Pricing | From about 0.20 yuan per image on Model Studio, or $0.035 per image on fal |
How to Use the Prompt Generator
Pick the model and the text load
Choose between the open-weight 2512 build and the API tiers, then say how much text belongs in the image. Type the exact wording into the text field, since that is what gets quoted and rendered verbatim.
Describe the scene in plain sentences
Give the subject, what is happening, the setting, the style and the lighting. Add the font style and medium if there is text, then put anything you do not want into the negative constraints field rather than the description.
Generate and paste
Copy the prompt into Qwen Chat, Alibaba Cloud Model Studio, fal, Replicate or a local ComfyUI workflow. If text comes out slightly wrong, edit that one quoted string rather than rewriting the whole prompt.
Frequently Asked Questions
Which Qwen image model should I use?
It depends on whether you need open weights. If you do, use Qwen-Image-2512 for generation and Qwen-Image-Edit-2511 for editing. Both are 20B, both are Apache 2.0, and both allow full commercial use. If you want the strongest quality and do not mind an API, Qwen-Image-3.0-Pro is the current flagship, released July 21, 2026. The important catch is that Qwen-Image-2.0 and 3.0 have never shipped weights, so anything describing 3.0 as open source is wrong. Also note there is no such thing as Qwen3-Image, and Plus and Max are API tier names rather than separate architectures.
Why does everyone say Qwen is the best at text in images?
Because for Chinese it genuinely was, by a wide margin. On LongText-Bench the original Qwen-Image scored 0.943 in English and 0.946 in Chinese, so it is almost language-symmetric. For comparison GPT Image 1 scored 0.956 in English but dropped to 0.619 in Chinese, and FLUX.1 Dev scored 0.005 in Chinese, which is effectively total failure. An independent team at HUST and ByteDance re-ran those numbers and reproduced them within a couple of points, so the claim holds up. What people miss is that English text was never the win: on CVTG-2K, Qwen-Image actually loses to GPT Image 1 on word accuracy.
Is Qwen still the best at text rendering in 2026?
No, and Alibaba's own benchmark says so. On Qwen-Image-Bench, which Alibaba built and published, the text rendering ranking puts GPT Image 2 first at 80.91, Seedream 5.0 second at 77.23, and Qwen-Image 2.0 Pro third at 72.74. Overall it places Qwen 2.0 Pro fifth. More awkwardly, Qwen-Image-2512 scored worse on text than the original Qwen-Image despite text being the selling point. The Chinese-language advantage is still real and still the reason to pick it, but the general lead is gone.
How do I make text render correctly?
Wrap every literal string in double quotes. Four separate official Qwen sources say this, including the prompt rewriter source code, which instructs that every piece of displayed text be enclosed in double quotation marks. Then give the surrounding detail in the order the rewriter expects: where the text sits, the exact transcription including punctuation and line breaks, the font style and colour and size, the medium it appears on such as chalkboard or neon or embroidery, its role such as title or slogan, its spatial relation to nearby objects, and finally the scene lighting as it affects legibility. Never write something vague like "a list of names", write the actual names.
What is the CJK trip-wire?
A quirk worth knowing because it silently ruins prompts. Qwen's rewriter detects the prompt language by scanning for any character in the CJK Unified Ideographs range. A single Chinese character anywhere routes your entire prompt through the Chinese rewriter, which rewrites the whole thing into Chinese. So an English prompt asking for a Chinese shop sign can come back fully translated. The safe pattern, which Qwen itself demonstrates, is to write the prompt body in one language and put the exact target glyphs verbatim inside quotes, since quoted text is explicitly protected from translation.
Should I write prose or comma-separated tags?
Prose, and this is not a style preference. The captions Qwen-Image was trained on were written to a spec requiring natural descriptive text without structured formats, and the rewriter tells the model to use fluent natural language in a single continuous block while strictly avoiding markdown lists, numbered items and headings. Writing tags puts you out of distribution. Guides recommending labelled fields like Subject: and Lighting: for Qwen are working against the documented training format.
How long should a Qwen-Image prompt be?
Around 200 words is the official recommendation, and that is what the rewriter targets. The hard limits are separate and worth knowing because the API truncates silently rather than erroring: 800 tokens for most models, 1,300 for the 2.0 series, and 4.5k for 3.0-Pro. Negative prompts cap at 500 characters.
Does Qwen-Image support negative prompts?
Yes, fully, on both the open weights and the API, which makes it unusual next to models like FLUX 3 that have none. Use it properly though: Qwen's own rewriter explicitly forbids negation words in the positive prompt, so "no chopsticks" should be handled by simply not mentioning chopsticks and putting the exclusion in the negative prompt field instead. Alibaba's own example negative prompt covers low resolution, deformed limbs and fingers, oversaturation, waxy skin, featureless faces, over-smoothing, an artificial AI look, chaotic composition, and blurry or distorted text.
What settings should I use for the open-weight models?
For Qwen-Image and Qwen-Image-2512, 50 inference steps with true_cfg_scale at 4.0. For Qwen-Image-Edit-2511, 40 steps at true_cfg_scale 4.0 with guidance_scale 1.0. When you are not using a negative prompt, pass a single space rather than an empty string, which is what the official examples do. Distilled variants exist if you want speed, including an 8-step LoRA. On memory, expect roughly 20.4 GB at fp8 and 40.9 GB at bf16, with quantized community builds fitting in about 13 GB.
How do I write editing instructions?
Be direct and specific, and always say what must stay the same. Vague instructions drift, so instead of "add an animal" write "add a light-gray cat in the bottom-right corner, sitting and facing the camera". For swaps the official phrasing is "replace Y with X" followed by X's key visual features. For anything involving a person, restate the identity details you want kept such as hairstyle, expression and clothing, because otherwise the face moves. Text changes are treated as replacements too, written as replace one quoted string with another. If you are combining a style change with other edits, put the style instruction last.
How do I reference multiple images?
You can pass one to three input images, and they are numbered by their order in the array, so your prompt references must line up one to one. Write them as Image 1, Image 2 and Image 3, or the Chinese equivalents. A working example reads like "the girl from Image 1 is wearing the black dress from Image 2 and sitting in the pose from Image 3". Input images should be between 384 and 2048 pixels per side and under 10 MB.
What is prompt_extend and should I leave it on?
It is Alibaba's automatic prompt rewriter and it defaults to on for every image endpoint. It only touches the positive prompt, never the negative one, and it adds a few seconds of latency. Qwen's own README strongly recommends leaving it on for editing, warning that results become unstable without it. Worth knowing that raw diffusers does not apply any rewriting unless you call it yourself, which is one reason local results can differ from the API for the same prompt.
What resolutions does Qwen-Image support?
The open-weight models use a fixed set: 1328 by 1328 for square, 1664 by 928 for 16:9, 928 by 1664 for 9:16, 1472 by 1104 for 4:3, 1104 by 1472 for 3:4, 1584 by 1056 for 3:2, and 1056 by 1584 for 2:3. The API models are more flexible, accepting any width and height as long as the total sits between 512 by 512 and 2048 by 2048, which is about 4.19 megapixels, with 2048 by 2048 as the default.
What does it cost and where can I run it?
Through Alibaba Cloud Model Studio, Qwen-Image-2.0 is about 0.20 yuan per image, 2.0-Pro about 0.50, and 3.0-Pro between 0.25 and 0.50 depending on output size. Note that 3.0-Pro is rate limited to 1 request per minute, so it is still effectively a preview. On fal it is $0.035 per image for qwen-image-2 and $0.075 for the pro tier. You can also use it free with rate limits in Qwen Chat, or run the open weights locally through ComfyUI, diffusers, Replicate or Together AI.
Is this tool free to use?
Yes. You get 3 free prompt generations per day with no signup required. For unlimited access, sign up for a Promptslove membership which includes all AI tools and 20,000+ premium prompts.
More image prompt generators
Want Unlimited AI Prompt Generation?
Get unlimited access to all AI tools, 20,000+ premium prompts, courses, and resources designed to maximize your creative output.
Qwen and Alibaba Cloud are trademarks of their respective owners. Promptslove is not affiliated with or endorsed by Alibaba. Model specifications, pricing and benchmark figures described here reflect published documentation at the time of writing and may change.