Qwen Image Prompts: A Working Guide to the Open Models

Qwen image prompts are written for a model that was trained to read layouts, not to match keywords. Qwen's own technical report describes a progressive training strategy that "starts with non-text-to-text rendering, evolves from simple to complex textual inputs, and gradually scales up to paragraph-level descriptions." That is the whole personality of the family. Give it exact strings, tell it where each one sits, and it will render them. Give it a comma-separated adjective pile and you get a generic image with garbled lettering.

What follows is a working qwen image prompt guide for the models you can actually run today: the prompt shape, the quirks that separate Qwen from Flux and Nano Banana (it takes real negative prompts, and on Alibaba's hosted API it rewrites your prompt by default), six copy-paste examples, and the editing dialect. If you want the release history, licensing, and where each version runs instead, that lives in our Qwen Image 3.0 explainer.

Which Qwen model your prompt is going to

This matters before a single word gets written, because the two branches of the family behave differently and only one of them is open.

• Qwen-Image-2512 is the current open-weight generation model, Apache 2.0, released December 31, 2025. Replicate describes it as "a 20 billion parameter text-to-image model with realistic human rendering, fine textures, and strong text rendering." This is the one behind most API access, the one in the ComfyUI ecosystem, and the one every prompt in this guide targets.
• Qwen-Image-Edit-2511 is the open editing branch, also Apache 2.0. Instruction plus image in, edited image out, with multi-image input supported.
• Qwen-Image-3.0 is the hosted flagship. It has no weights, no API outside an invite gate, and no published prompting guide, so the techniques below are written for the open line. What 3.0 changes is covered further down.

One disclosure up front, because it changes how to read what follows. Dream Pixel Forge does not ship a Qwen image model today. We run Nano Banana, GPT Image 2, and Grok Imagine, plus FLUX.1 Kontext Pro for edits. Qwen-Image-3.0 is on the list to add once its access opens. The techniques here come from Qwen's repository and technical report, the Hugging Face model cards, Alibaba Cloud's Model Studio documentation, and Qwen's published demos, not from Qwen running in our own product. The renders under the example prompts are ours: we ran those prompts through qwen/qwen-image-2512 on Replicate so you can see what the words in this guide actually produce, misses included.

Where you run Qwen also changes which advice applies, because the surfaces differ in what they expose. Local weights give you every parameter. Replicate exposes most of them. Alibaba Cloud's Model Studio adds a prompt rewriter that is on by default. Qwen Chat gives you a text box and little else. Each section below says which path it is talking about.

How to prompt Qwen Image: the shape that works

Qwen reads sentences. The reliable structure, drawn from the example prompts Qwen ships in its own repository and its launch write-up, is five slots in this order:

1. Artifact type. "A letterpress poster", "a chalkboard menu board", "an editorial infographic". Naming the object you are making is the highest-value phrase in the prompt, because it tells the model which layout grammar to use.
2. Scene and subject in complete sentences.
3. Every text string, quoted and placed. Not "with a sign". The literal characters, plus where they sit relative to something else in the frame.
4. Typography and color for each string: weight, case, serif or sans, color.
5. Camera, light, and finish last.

Qwen's own canonical demo prompt is the clearest illustration of slot three doing the work, and it is worth reading closely because almost nobody writes prompts this specific:

A coffee shop entrance features a chalkboard sign reading "Qwen Coffee
$2 per cup," with a neon light beside it displaying "通义千问". Next to
it hangs a poster showing a beautiful Chinese woman, and beneath the
poster is written "π≈3.1415926-53589793-23846264-33832795-02384197".

Three separate text surfaces, each with its literal content and its physical placement, in two writing systems, plus a numeric string with no semantic meaning to lean on. That is the job Qwen was built for, and the prompt spends nearly all of its words on it.

One more habit from the repo: Qwen appends a fixed quality suffix to English prompts in its own reference implementation, , Ultra HD, 4K, cinematic composition., with a Chinese equivalent. It is a small, free improvement on photographic work, and Qwen calls the pair "positive magic" in the code. Skip it on flat graphic and print work, where "cinematic" drags the render toward depth of field and warm grading you did not ask for.

Qwen takes real negative prompts, if your path exposes the field

This is the biggest dialect difference in the family. FLUX.2 has no negative prompt field at all, and on Nano Banana, GPT Image 2, and Grok Imagine we consistently get better results describing the state we want than the state we do not. Qwen is the exception: negative_prompt is a real, weighted input on both the open weights and the Alibaba Cloud API, where Alibaba documents it as accepting Chinese or English up to 500 characters.

The caveat is that the field is not universal. Running the weights locally exposes it, and so does Replicate, but Qwen Chat and most hosted wrappers do not. So read this section as a technique for local and raw-API work, and on any path without the field, fold the same intent into the positive prompt: describe sharp, cleanly kerned lettering rather than asking for the absence of blurry text.

Qwen's own default in its 2512 example code is a quality laundry list rather than a content filter:

低分辨率，低画质，肢体畸形，手指畸形，画面过饱和，蜡像感，
人脸无细节，过度光滑，画面具有AI感。构图混乱。文字模糊，扭曲。

Translated, that is low resolution, low quality, deformed limbs, deformed fingers, oversaturated, wax-figure look, faceless detail, excessive smoothness, an AI look, chaotic composition, and blurry or distorted text. Two things follow from it. First, this is the correct use of a Qwen negative: broad failure modes, not scene content. Asking a negative to remove an object from a described scene works badly; asking it to suppress a rendering artifact works well. Second, note the last clause. "Blurry, distorted text" in the negative is Qwen's own recommendation, and it is the one negative term that earns its place in almost every text-heavy prompt.

When you have nothing to suppress, do not leave the field empty. Qwen's reference code passes a single space and comments it as the recommended value: negative_prompt = " ". Our general position on negatives across models has not changed, and the reasoning is in negative prompts. Qwen is the model where the exception is real, and it is narrower than most people use it.

On some paths a language model rewrites your prompt first

Qwen ships an official prompt enhancement tool, rewrite() in tools/prompt_utils.py, powered by Qwen-Plus, and a matching polish_edit_prompt() for the edit model powered by Qwen-VL-Max. The repo recommends it. On Alibaba Cloud's hosted API the same idea is a parameter, prompt_extend, and Alibaba's documentation is explicit that it defaults to on: "Enable smart prompt rewriting. The model optimizes the positive prompt."

Understand what that means for control. A one-line brief gets silently expanded into a long, styled paragraph before the image model ever sees it, and the expansion invents lighting, lens, and mood. That is genuinely useful when you are exploring and actively hostile when you are matching a brand or reproducing a layout you have already approved. The rule we hold to:

• Exploring? Leave rewriting on and write short. It is a better idea generator than a blank page.
• Producing? Turn it off, and write the detail the rewriter would have added. A rewritten prompt is not reproducible, and reproducibility is most of what production work needs.
• Debugging a bad render? Turn it off first, before you change any words. Half the time the prompt you are blaming is not the prompt the model received.

Know which path you are on before you go looking for the switch, because it does not exist everywhere. prompt_extend is specific to Alibaba's Model Studio API and is not part of the Replicate surface at all. On a path that does not rewrite, the words you wrote are the words the model reads, which cuts both ways: results are predictable and reproducible, and a short lazy prompt stays short and lazy rather than being quietly padded into something better.

Prompt length is the other thing to check against your access path. Alibaba documents 1,300 tokens for the qwen-image-2.0 series and 800 tokens for the rest, which is a real ceiling on the hosted API and much shorter than the 4.5k figure attached to 3.0.

Qwen prompts worth copying

Six prompts for jobs people actually have. Each one is complete and copy-pasteable, and the note underneath names the lever doing the work so you can adapt rather than paste.

A poster with an exact headline

A vintage letterpress event poster printed on textured cream stock,
shot flat and square to the camera. Across the upper third, the words
"NIGHT MARKET" in large condensed sans-serif capitals in deep red.
Centred beneath it, the line "Fridays 6pm until late" in a small
italic serif in charcoal. Along the bottom edge, "Pier 9, East Dock"
in widely letterspaced capitals. Flat two-colour print, visible paper
grain, subtle ink misregistration, even overhead light.

Three strings, three positions, three type treatments. The prompt never says "with some text about a market", which is the phrasing that produces decorative squiggles on every model in this class.

Full size

A bilingual menu board, which is Qwen's home ground

A hand-lettered chalkboard menu board mounted on a whitewashed brick
wall in a small cafe. The header reads "TODAY'S POUR" in chalk capitals
with "今日推荐" directly beneath it in slightly smaller Chinese
characters. Below the header, three left-aligned lines with prices
right-aligned: "Ethiopia Guji  4.50", "House Filter  3.20", and
"Cold Brew  4.00". Warm tungsten light from the left, shallow depth of
field, photographic, sharp focus on the board.

The multilingual claim is the reason to reach for Qwen at all, and the research paper is specific about it: the model "performs exceptionally well in alphabetic languages such as English, but also achieves remarkable progress on more challenging logographic languages like Chinese." Note what makes this prompt work beyond the character strings: it states the alignment. Left-aligned items with right-aligned prices is layout instruction, and layout instruction is where the long-prompt capability pays.

Full size

A process infographic

An editorial infographic titled "How Cold Brew Works", laid out as four
numbered steps in a single horizontal row on a warm off-white
background. The four step labels read "1. Coarse grind", "2. Steep 16
hours", "3. Filter twice", and "4. Dilute to taste", each above a
simple line-art icon. Thin dark grey line work with one amber accent
colour, all text in a bold geometric sans-serif, generous white space,
no photographic elements, flat vector style.

Five strings and an explicit arrangement. Naming the layout ("a single horizontal row") is what stops the model from inventing a collage, and it is the instruction most infographic prompts omit.

Full size

A product shot with legible label copy

A commercial product photograph of a matte black cold brew coffee can
standing upright on a wet slate slab. On the front of the can, the
label reads "MIDNIGHT ROAST" in heavy condensed sans-serif capitals,
with the smaller line "cold brew, no sugar" beneath it in thin
uppercase letters. Studio softbox lighting from above and slightly
right, one soft shadow falling left, 85mm lens, shallow depth of field.
Plain deep charcoal background, edge to edge, photorealistic.

If your path exposes the field, the negative prompt for this one is blurry text, distorted text, oversaturated, wax-figure appearance, chaotic composition. If it does not, the positive prompt above already carries the load. The explicit background statement is the single most commonly omitted line in AI product photography prompts on any model, and without it Qwen invents a set dressing that has to be cropped out later.

Full size

A portrait that uses what 2512 improved

A candid indoor photograph of a woman in her early thirties standing at
a crowded weekend market stall, holding a paper cup and half turned
toward the camera mid-laugh. Visible skin texture and fine pores, loose
dark hair, no retouching. Ordinary overhead market lighting, no staged
studio light, the unpretentious framing of a phone snapshot, natural
colour, sharp focus on her face.

Qwen's 2512 release notes lead on "Enhanced Human Realism" and say the update "significantly reduces the AI-generated look." The way to use that is to describe ordinary light and casual framing, which is exactly what Qwen's own demo prompt for this release does when it asks for an image that "resembles a casual iPhone snapshot: unpretentious composition." Studio lighting language pulls the render back toward the glossy default you were trying to escape.

Full size

An edit instruction, which is a different dialect

Replace the text on the shop awning so it reads "FERN & CO" in the same
painted serif lettering, matching the existing colour, weathering, and
perspective. Keep the awning shape, the storefront, the people on the
pavement, the lighting, and the composition exactly the same.

Name the change, then protect everything else by listing it. This is covered properly in the next section, but the shape is visible here: one instruction, one explicit preservation clause, no re-description of the scene.

Editing prompts: change one thing, protect the rest

Qwen-Image-Edit handles, in Qwen's own list, "style transfer, additions, deletions, detail enhancement, text editing, and character pose adjustment", and the architecture is built around not destroying the source: the report describes a dual-encoding mechanism feeding the input image to both Qwen2.5-VL and a VAE encoder to "balance between preserving semantic consistency and maintaining visual fidelity."

Four rules make edits behave.

• Name only what changes. Re-describing the whole frame is a regeneration wearing an edit's clothes, and it degrades everything that was already right.
• List what must survive. Composition, lighting, faces, background. Anything unprotected is fair game for reinterpretation.
• Avoid pronouns. Write "the woman in the red coat", not "her". Referring expressions resolve badly across every editing model we run.
• Use the edit rewriter. Qwen recommends prompt rewriting for edit stability specifically, via polish_edit_prompt(), which reads the input image with Qwen-VL-Max before rewriting. Editing is the one job where we would leave a rewriter on, because it reads the source image rather than guessing.

The 2511 edit release targets the failure people complain about most: Qwen lists image-drift mitigation and character consistency first among its changes. Drift is the slow degradation you get when you edit an edit of an edit, and the practical defence is structural rather than lexical. Always edit from the original, never from the third generation of a chain. Holding one identity across a whole set is a separate problem that no prompt solves, and we cover why in consistent AI characters.

Settings that change the output as much as words do

If you are running the weights yourself, three numbers from Qwen's own example code matter more than another adjective. true_cfg_scale at 4.0 is the documented starting point, with num_inference_steps at 50 for generation and 40 for edits. Higher guidance buys literal prompt adherence and costs naturalism, which is the trade every diffusion model makes.

If you are calling it through a hosted service you mostly inherit those choices. Replicate's defaults for this model are guidance 4 and 40 steps with its fast sampling path on, which is the usual quality-for-latency trade, and no fixed seed unless you pass one. That last one is worth planning around: without a pinned seed the same prompt gives you a different image every run, so an approved render is a file to keep rather than a recipe to re-run. Save the output, not just the words.

Resolution is the other one. The open 2512 model ships a fixed aspect-ratio table rather than free-form dimensions, and picking from it avoids the soft, stretched output you get by inventing a size. These are the dimensions Qwen's own reference implementation pairs with each ratio:

Hosted services usually expose the ratio rather than the pixel dimensions, so treat the second column as the shape you are asking for rather than a guarantee. On Replicate a 16:9 request comes back as a 1664 x 928 webp, which matches the table. That long-edge ceiling is the practical limit of the open line, and it is the honest reason to send a print job elsewhere.

One more expectation to set: a 20B model on a rented GPU is not instant. Generation itself runs in about ten seconds, but a cold start on a serverless host pushes the wall-clock wait past a minute. That is a fine trade when Qwen is the right model for a text-heavy layout and an irritating one when you are just exploring, which is another argument for settling your copy before you start rendering.

Worth saying plainly: the layout-first habit this guide teaches is not Qwen-specific. Naming the artifact type, quoting every string, anchoring each one to a position, and describing the typography is the technique that separates good text rendering from garbled text rendering on every model that can render text at all. If you want to practise it without setting up weights or a second billing account, run the same prompts against Nano Banana Pro or GPT Image 2, which are the two strongest text models we operate, and you will see the same rules bite.

What Qwen-Image-3.0 adds, and why these techniques still apply

Qwen announced Qwen-Image-3.0 in July 2026 under the single word "Real", positioning it as a step past 1.0's "Precision" and 2.0's longer list. The announcement claims a 4.5k-token prompt window, text legible down to 10 pixels, native rendering across 12 languages, and over 100 artistic styles. The showcase is a 3x3 grid of nine separate dense infographics generated in one pass from a 3.7k-token prompt, plus a nested-UI demo rendering a VSCode window containing a Qwen Chat window containing a WeChat window containing a coffee poster.

3.0 is currently reachable only through Qwen Chat, with no published weights, no model card, no technical report, and API access behind an invite gate, so it is not where your prompts go today and there is no official 3.0 prompting guide to follow. What its demos confirm is the rule this whole guide runs on: every capability Qwen shows off at 3.7k tokens is structural, not stylistic. If your job is dense multilingual layout and 2512 is not holding it, waiting on 3.0 access is rational. For everything else, the open model plus a better-structured prompt closes most of the gap.

Qwen-Image-3.0 is on the list to add here, and the fastest way to get it the day it lands is to have an account already: create one and the new engine shows up in the same model dropdown, on the same credits, with no migration. Until then the studio runs Nano Banana, GPT Image 2, and Grok Imagine on one balance, so the prompt habits above have somewhere to go in the meantime.

Fixing a bad Qwen render: one change at a time

Rewriting the whole prompt destroys the information about what was already working. Match the symptom to a single change, re-run, then move on.

Where Qwen sits against the other models

Prompt craft does not transfer between families, and treating one dialect as universal is the most expensive habit in this field. Qwen rewards structure and exact strings. Nano Banana rewards literal declarative sentences and leads at reference-driven editing. GPT Image 2 rewards naming the artifact type first and has the tightest content filter. Grok Imagine rewards camera and lens vocabulary and is weak at legible text. Flux rewards prose and a single lighting clause. The wider field is mapped in best AI image generators, and AI art styles covers around thirty style registers that work on any of them.

The bottom line

Write Qwen prompts like a layout brief rather than a mood board. Name the artifact type first, quote every string exactly and anchor it to a physical position, describe the typography, and spend your remaining words on light rather than adjectives. Pick an aspect ratio from the supported table. Then check what your particular path gives you: if there is a negative-prompt field, keep it to rendering failures rather than scene content, and if there is a rewriter, turn it off before you judge a result, because you cannot debug a prompt you did not send.

Do that and Qwen does the thing it is genuinely best at in the open-weight field: dense, multilingual, text-heavy layouts that other models turn into decorative nonsense. Ask it for a pretty picture with a vague vibe and you will wonder what the fuss was about.