Qwen Image 3 Prompts: 50 Copy-Paste Ideas [2026]

Qwen-Image-3.0 is Alibaba's third-generation image model, released July 21, 2026. Alibaba claims it renders literal text down to roughly 10px across 12 languages and 20+ fonts, accepts prompts up to about 4,500 tokens, and handles text-to-image plus instruction editing in one model. The technique that unlocks all of it is small: put the exact words you want rendered inside quotation marks, then describe where those words sit relative to each other. On Morphed it costs 12 credits per image in both generate and edit mode, with output up to 2K.

Below are 50 copy-paste prompts built around that technique, split by the job they do. Every one of them was written to be pasted without edits, then adapted by swapping the quoted strings.

Which Qwen Model Are You Actually Using?

Three separate things share the Qwen name, and mixing them up is the fastest way to follow advice that cannot work. Qwen3 is a text large language model and produces no images. Qwen-Image-3.0 is the current image model. The older Qwen-Image line is a different set of models with different licensing.

The important distinction: Qwen-Image-3.0 launched with a marketing gallery and nothing else. No weights, no license, no model card, no technical report, and no benchmark scores. That broke Alibaba's own open-release pattern and became the main story in the coverage around launch. Treat every capability claim in this guide as a vendor claim confirmed only by hands-on output, not by a published benchmark.

The practical consequence for prompting: any tutorial you find about LoRA training, ComfyUI nodes, or self-hosting a Qwen image model is describing the older Apache 2.0 line. None of it applies to 3.0, which is hosted-access only. If you want fine-tuned control over a character or a brand style, that workflow lives on the open models or on a different platform entirely.

Why Quoting the Exact Text Changes the Output

Most weak text prompts describe the text instead of supplying it. The model then has to invent the words, and invented words are where garbled lettering comes from. Supplying the literal string in quotation marks converts the task from creative writing to typesetting, which is what this model is actually good at.

Here is the same poster brief written both ways.

Weak: "A poster for a jazz festival with the name and dates on it in a cool retro font, plus the venue at the bottom."

Strong: "Vintage screen-printed jazz festival poster on off-white paper stock with visible halftone texture. Large bold condensed capitals across the upper third reading "BLUE HOUR". Directly beneath the headline, small widely-spaced capitals reading "SEPTEMBER 12 TO 14". Along the bottom edge, one line of small caps reading "PIER 7, OAKLAND". Muted teal and burnt orange, two-color risograph look. No other text."

Five rules do the work:

1. Quote the literal string. Anything you want rendered goes inside quotation marks exactly as it should appear, including punctuation and capitalization.
2. Position by relationship, never coordinates. "Across the upper third", "directly beneath", "along the bottom edge", "in the lower right corner". Pixel values and percentages are ignored.
3. Build hierarchy in words. "Large bold condensed capitals" against "small widely-spaced capitals" tells the model the relative weight. Point sizes do not.
4. Paste real characters for non-Latin scripts. Write 秋の収穫祭, not "the Japanese words for autumn harvest festival". Describing a script in English produces decorative nonsense.
5. End with a text lock. Adding "no other text" stops the model from filling empty margins with invented lettering, which is the single most common failure on poster layouts.

Render at the highest resolution available. Small type is where this model earns its reputation, and 2K output gives caption-sized text roughly four times the pixel area of a 1K render to resolve into.

Poster Prompts With Headline, Subhead, and Body Copy

Posters are the clearest test of the quoting method because they need three distinct type sizes in one image. Each of these specifies all three.

Prompt: "Swiss modernist concert poster, deep red background, thin white grid. Large bold sans-serif capitals across the upper third reading "NIGHT SIGNAL". Directly beneath in small light capitals: "A LIVE ELECTRONIC PERFORMANCE". Bottom fifth carries three lines of small body copy reading "FRIDAY 19 SEPTEMBER", "DOORS 8PM", "UNION HALL, MANCHESTER". Generous margins. No other text."

Prompt: "Minimal film festival poster, matte black background with a single beam of warm light across the center. Centered thin serif capitals reading "THE LONG RETURN". Small italic subhead directly beneath reading "Selected Works 2019 to 2026". Along the bottom, a dense five-line credit block in tiny grey type, legible but small. No other text."

Prompt: "Retro travel poster, flat vector mountains in dusty orange and cream, mid-century printing texture. Arched capitals across the top reading "VISIT THE HIGH BASIN". Below the illustration, a two-line block reading "NATIONAL PARK SERVICE" and "ESTABLISHED 1934". Small route information in the lower left corner. No other text."

Prompt: "Bold protest-style poster, high-contrast black on yellow, thick condensed woodblock type. Headline stacked across three lines reading "READ", "THE", "FINE PRINT". Small dense paragraph of body copy in the lower third, five lines of legible small type. Rough ink edges. No other text."

Prompt: "Editorial art exhibition poster, cool grey paper, single duotone photograph of a folded sheet of steel occupying the left half. Right column holds a headline in large medium-weight capitals reading "SURFACE TENSION", a subhead reading "New sculpture by Ana Ferreira", and a five-line information block with dates, venue, and opening hours in small type. No other text."

Prompt: "Vintage boxing match poster, cream paper with heavy aging and torn edges. Enormous stacked capitals reading "TUESDAY NIGHT", then "MAIN EVENT". Two fighter names in equal-weight capitals separated by a small "VS". Bottom banner reading "CIVIC ARENA, DOORS 7PM". Ornate Victorian display faces mixed with slab serif. No other text."

Storefront and Signage Prompts That Keep the Lettering Clean

Signage scenes are harder than posters because the text has to sit on a surface at an angle and still read correctly. Name the sign material, because material dictates letterform quality.

Prompt: "Photograph of a narrow corner bakery at dawn, wet pavement, warm interior light. Hand-painted gold-leaf lettering on the front window reading "MORNING & CO." with smaller lettering beneath reading "BREAD, PASTRY, COFFEE". Shallow depth of field, 35mm, natural light. No other text."

Prompt: "Night photograph of a Tokyo backstreet ramen bar, red noren curtain, condensation on the glass. Vertical illuminated sign reading 一番らーめん. Small handwritten menu board beside the door listing four items in Japanese. Cinematic low light, 50mm. No other text."

Prompt: "Split-flap departure board inside a 1970s train station, warm tungsten light, slight motion blur on one flipping row. Six rows of destinations and times, all precisely aligned in columns headed "TIME", "DESTINATION", "PLATFORM". Mechanical character shapes, correct column spacing. No other text."

Prompt: "Documentary photograph of an outdoor farmers market stall, overcast light. Three hand-lettered chalkboard signs propped among produce reading "HEIRLOOM TOMATOES £3/KG", "COLD PRESSED CIDER", and "NO PLASTIC BAGS PLEASE". Chalk texture, slightly uneven letterforms. No other text."

Prompt: "Wide shot of an independent bookshop facade in Lisbon, azulejo tiles, late afternoon sun. Painted wooden fascia sign reading "LIVRARIA DO MONTE" in worn cream serif capitals. A small window card reading "ABERTO" hangs at eye level. Warm film tones. No other text."

Infographic and Dense Single-Page Layout Prompts

This is the category the model was built to show off. A 4,500 token prompt ceiling means you can specify a full page structure in one pass rather than assembling panels afterward. The trade is verification time: a dense page has twenty places for a typo to hide.

Prompt: "Front page of a broadsheet newspaper, aged newsprint texture. Masthead across the top reading "THE HARBOUR REVIEW". Below it a dateline strip, then a five-column grid: one lead story with a large headline reading "TIDE GATES CLOSE EARLY AS STORM TURNS NORTH", one black and white photograph with a caption, three secondary headlines, and a boxed weather panel in the lower right. Body copy is legible small type throughout. No other text."

Prompt: "Clean flat-design infographic poster titled "HOW A HEAT PUMP MOVES ENERGY". Four numbered stages arranged left to right, each with a simple line icon, a short bold label, and two lines of explanatory body copy beneath. A horizontal arrow connects the stages. Muted blue and warm grey palette, generous white space, thin rules separating sections. No other text."

Prompt: "Single-page conference schedule, modern editorial layout. Header reading "SIGNALS 2026" with a subhead reading "Day One, Main Stage". Below, a two-column time grid with eight rows: left column holds times in small bold type, right column holds session titles and speaker names in regular weight. Alternating row shading. Footer with venue address. No other text."

Prompt: "Vintage botanical study sheet on aged paper, single fern rendered in fine ink detail at the center. Handwritten-style annotations arranged around the specimen with thin leader lines pointing to the frond, stem, and spore cluster. A title block in the lower right corner reading "POLYSTICHUM SETIFERUM" with three lines of collection notes beneath. No other text."

Prompt: "Storyboard sheet for a commercial, six numbered panels in a two by three grid on white. Each panel contains a rough greyscale sketch of a scene, a shot label reading "WIDE", "CLOSE", or "TRACKING", and two lines of action description beneath the frame. Header strip reading "SPOT 01 / 30 SECONDS". Clean production-document look. No other text."

Prompt: "Printed restaurant menu photographed flat on a dark oak table, warm side light. Header reading "CASA VERDE" with a small subhead reading "Kitchen open until 11". Three sections headed "TO START", "MAINS", and "TO FINISH", each listing four dishes with a short description line and a price aligned to the right margin. Letterpress texture on heavy card. No other text."

Exploded Product Diagram Prompts With Labeled Callouts

Exploded diagrams stress two things at once: spatial coherence between separated parts and the leader lines connecting labels to the right components. State the separation axis and the label style explicitly.

Prompt: "Technical exploded diagram of a running shoe, components separated vertically along a single axis: outsole, midsole foam, insole, upper mesh, laces. Thin leader lines connect each part to a small label on the right margin reading "RUBBER OUTSOLE", "COMPRESSION FOAM", "REMOVABLE INSOLE", "ENGINEERED MESH", and "FLAT LACE". Clean white background, soft studio shadow, product manual style. No other text."

Prompt: "Patent-style line drawing of a mechanical wristwatch, exploded along the horizontal axis, black ink on cream paper. Each component carries a small numbered callout with a matching legend list in the lower left corner. Precise hairline weights, no shading, engineering drawing conventions. No other text."

Prompt: "Exploded view of a wireless earbud charging case in matte white, parts floating apart with even spacing, isometric angle. Callout labels in small grey sans-serif capitals connected by thin leader lines: "HINGE ASSEMBLY", "MAGNET RING", "BATTERY CELL", "USB-C BOARD". Soft gradient background. No other text."

Prompt: "Cutaway cross-section illustration of an espresso machine, front half removed to reveal the boiler, pump, group head, and water reservoir. Each internal component labeled with a short text tag connected by a leader line. Flat vector style, three-color palette of chrome grey, deep red, and cream. No other text."

Prompt: "Assembly instruction sheet for a flat-pack chair, exploded isometric drawing at the center, parts numbered one through nine. A parts list runs down the left margin with quantities and short names. A small tools panel in the lower right shows an allen key and screwdriver. Line art only on white. No other text."

Quote Card and Social Graphic Prompts

Social graphics fail when the quoted text runs long and wraps badly. Specify the line breaks yourself by writing the string as separate stacked lines.

Prompt: "Square social graphic, deep navy background with a fine noise texture. Centered quotation in large medium-weight serif, broken across three lines reading "Craft is what", "remains after", "the deadline passes". Small attribution beneath in widely-spaced capitals reading "PRODUCTION NOTES". Generous margins. No other text."

Prompt: "Bold Instagram carousel cover, flat coral background. Huge condensed capitals filling the frame across four stacked lines reading "STOP", "GUESSING", "YOUR", "PRICING". A thin white rule beneath, then small capitals reading "SWIPE FOR THE MATH". No other text."

Prompt: "LinkedIn post graphic, clean white background with a single thin left border rule in emerald. Headline in bold sans-serif reading "Three things I got wrong about hiring". Beneath it a numbered list of three short lines in regular weight. Small circular avatar and a name line in the bottom left corner. No other text."

Prompt: "Minimal typographic quote card on textured cream paper, single line of elegant italic serif centered vertically reading "Measure twice, ship once". A small pressed-ink ornament above the line. Warm soft shadow as if photographed. No other text."

Prompt: "Dark-mode statistic card for social sharing, charcoal background. Enormous numeral reading "68%" occupying the upper half in bright lime. Beneath it, two lines of white body copy reading "of teams reported faster review cycles" and "after moving to weekly releases". Small source line in grey at the bottom. No other text."

Bilingual and Multilingual Layout Prompts

Paste the real characters. Then state how the two scripts pair, because the model otherwise treats the second script as decoration and sizes it wrong.

Prompt: "Japanese autumn festival poster, deep indigo background with a rising moon and scattered maple leaves. Large vertical brush-style characters down the right side reading 秋の収穫祭. Paired beneath in small horizontal Latin capitals reading "AUTUMN HARVEST FESTIVAL". The Japanese is the dominant element, the English sits as a quiet secondary line. Small date block in the lower left. No other text."

Prompt: "Modern civic poster for a city library, warm paper stock, flat two-color illustration of stacked books. Large bold characters across the upper third reading 城市图书馆. Directly beneath in medium-weight Latin capitals reading "CITY LIBRARY". Both lines share the same optical weight and the same left margin. Three lines of small opening-hours copy at the bottom. No other text."

Prompt: "Korean coffee festival poster, cream background with a hand-drawn cup illustration. Headline in bold Korean characters reading 서울 커피 페스티벌. Small Latin subhead directly beneath reading "SEOUL COFFEE FESTIVAL". Dates in the lower right. Clean editorial spacing. No other text."

Prompt: "Bilingual museum wall label photographed straight on, brushed aluminium plate. Top block in English reading "Study for a Standing Figure", then "Bronze, 1962". Lower block repeats the information in French reading "Étude pour une figure debout" and "Bronze, 1962". English is bold, French is lighter and slightly smaller. Etched lettering. No other text."

Prompt: "Airport wayfinding sign photographed at an angle, backlit white panel, standard transport typeface. Primary line in large capitals reading "BAGGAGE RECLAIM", secondary line beneath in Arabic script, and a large directional arrow to the right. Clean sans-serif, high contrast, correct optical spacing between scripts. No other text."

A caution that belongs with this section: independent testers checking Korean output character by character found mixed-up vowels and misspelled words inside layouts that looked entirely clean at a glance. Support for 12 languages and correct rendering in 12 languages are different claims. Have a native reader check every character before anything with non-Latin script ships.

UI Mockup Prompts for App Screens and Dashboards

UI mockups need component vocabulary, not visual adjectives. Name the elements a designer would name.

Prompt: "Mobile app screen mockup on a clean white background, iOS style. Top navigation bar with a back chevron and a centered title reading "Portfolio". Below it a balance card showing a large figure reading "£12,480.20" with a small green change indicator reading "+2.4% today". Beneath, a list of four holdings, each row showing a ticker, a name, and a value aligned right. Bottom tab bar with four icons. No other text."

Prompt: "Desktop SaaS dashboard mockup, light theme, 16:9. Left sidebar with six navigation items reading "Overview", "Projects", "Team", "Billing", "Reports", and "Settings". Main area holds a page title reading "Overview", four metric cards in a row, and a line chart panel beneath with a legend. Clean spacing, subtle borders, no drop shadows. No other text."

Prompt: "Dark-mode music app screen, deep charcoal background. Album artwork square at the top, track title beneath reading "Slow Interference", artist line reading "Hale Court". Scrub bar with timestamps reading "1:42" and "-2:18". Playback controls beneath. Small queue list at the bottom with three upcoming tracks. No other text."

Prompt: "Landing page hero mockup shown in a browser chrome frame, address bar reading "northloop.io". Headline in large bold sans-serif reading "Ship your API in a weekend". Subhead beneath reading "Managed infrastructure for small teams". Two buttons side by side reading "Start building" and "Read the docs". Clean product screenshot below the fold line. No other text."

Prompt: "Checkout flow screen mockup, light theme, single centered column. Step indicator at the top with three labels reading "Cart", "Details", and "Payment", with the second step active. Form fields labeled "Full name", "Email", and "Address". A summary panel on the right lists two line items and a total reading "$148.00". Primary button reading "Continue to payment". No other text."

Editorial Scenes With Multiple Subjects

Multi-subject scenes are where photographic detail matters more than typography. Assign each person a distinct role and position so the model does not blend them.

Prompt: "Editorial photograph of a four-person kitchen brigade during service, shot from the pass. Head chef in the center plating, two cooks at the stations behind, one runner blurred in the foreground carrying plates. Steam, tungsten overhead light mixing with cool daylight from a high window. Visible skin texture, candid expressions, 35mm reportage style."

Prompt: "Magazine spread photograph of three architects reviewing drawings on a large table, natural window light from the left. One standing and pointing, two seated. Rolled plans, scale models, and coffee cups on the surface. Muted palette, shallow depth of field on the foreground model, sharp on the standing figure."

Prompt: "Documentary frame of a small print workshop, two people operating a letterpress. Older printer feeding paper, younger apprentice inking the plate. Ink-stained apron detail, warm bulb lighting, dust in the air. Fine texture on paper stock and metal type. 50mm, natural grain."

Prompt: "Wide editorial portrait of a five-person startup team standing in a converted warehouse office, exposed brick and steel beams. Staggered depth so no two faces overlap, each person in different casual clothing, natural expressions rather than posed smiles. Overcast daylight through tall windows, soft contrast."

How Edit Mode Handles Reference Images

Edit mode on the Qwen Image 3 model page takes up to three reference images and costs the same 12 credits as generation. Three rules govern whether an edit lands.

Image order is the addressing system. The model numbers references in upload order. Image 1 is your first upload. If you reorder uploads and leave the prompt unchanged, the instruction silently points at the wrong source. Restate the mapping in the prompt when the edit is complex.

Write substitutions, not descriptions. "The woman from Image 1 wearing the black coat from Image 2" gives the model a source and a target. "A woman in a black coat" throws away both references and regenerates from scratch.

Use negative prompts for the known failure modes. Short negative strings like "blur, extra fingers, duplicated text" cost nothing and catch the artifacts that appear most often in edits.

Prompt: "The woman from Image 1 is wearing the black wool coat from Image 2, standing in the same pose as Image 1. Keep her face, hair, and the background unchanged. Negative prompt: blur, extra fingers, warped hands."

Prompt: "Replace the text on the shop sign in Image 1 so it reads "THE PAPER MILL". Match the existing font, curvature, weathering, and lighting exactly. Change nothing else in the frame."

Prompt: "Place the product bottle from Image 2 onto the marble surface in Image 1. Match the direction and softness of the existing light, add a contact shadow beneath the bottle, and keep the reflection consistent with the surface. Negative prompt: floating object, mismatched shadow."

Prompt: "Take the room from Image 1 and restyle it in the material palette of Image 2. Preserve the exact camera angle, window positions, and furniture layout. Only surfaces, colors, and materials should change."

Prompt: "The person in Image 1 is holding the coffee cup from Image 2 in their right hand. Adjust the hand position so the grip is anatomically correct. Keep the face and clothing identical. Negative prompt: extra fingers, merged hands, duplicate cup."

Prompt: "Correct the spelling on the poster in Image 1. The headline should read exactly "SUMMER SESSIONS". Preserve the existing letterform style, kerning rhythm, color, and paper texture. Change no other element."

Prompt: "Extend the image in Image 1 to a 16:9 landscape crop. Continue the wall texture, floor line, and lighting gradient naturally into the new area. Do not add new objects or people."

Prompt: "Combine three sources: the model from Image 1, the dress from Image 2, and the studio backdrop and lighting setup from Image 3. The model keeps her face and hair from Image 1 and her pose from Image 1. Negative prompt: blur, distorted proportions, seam artifacts."

Prompt: "Remove the two parked cars from the street scene in Image 1 and reconstruct the road surface, kerb line, and building base behind them. Keep the lighting, shadows, and every sign legible and unchanged. Negative prompt: smudged texture, warped architecture."

What We Found Testing Text-Heavy Scenes on Morphed

These observations come from running text-dense prompts through Qwen Image 3 on Morphed rather than from a published benchmark, since no benchmark for this model exists. They are qualitative and worth reading as such.

Four results held up well. A farmers market scene with three separate hand-lettered signs came back with all three spelled correctly and each one sitting on its board at a plausible angle, which is the case where most models start inventing a fourth sign. A split-flap station board produced precise rows with columns that stayed aligned down the whole board rather than drifting. A request for gold-leaf café window lettering returned the exact string specified, with the leaf catching light along the letter edges. An exploded running-shoe diagram kept the parts, the labels, and the leader lines mutually coherent, meaning each line actually terminated at the component its label named.

Two failure modes showed up repeatedly. Exact object counts drift: ask for six items and you may get five or seven, laid out so convincingly that the miscount is easy to miss on first review. Geographic layouts are worse. Maps, floor plans, and anything with real-world spatial relationships come back looking authoritative and measuring wrong, which is the most dangerous kind of error because the output invites trust.

The workflow that follows from this: generate at 2K, zoom to 100 percent, and read every rendered character before the file leaves your machine. On a dense infographic that check takes about as long as writing the prompt did. Budget for it. At 12 credits per image, a four-pass typography refinement costs 48 credits, so it is cheaper to spend the extra two minutes specifying exact strings up front than to iterate blind.

When Qwen Image 3 Is the Wrong Choice

Three situations where reaching for this model costs you time.

People-first photoreal work. When the shot lives or dies on skin texture, micro-expression, and lens character, a FLUX or Nano Banana style model gets there in fewer attempts. Qwen Image 3 renders convincing material and skin detail, but the models tuned for portraiture still lead on faces. Start with the Nano Banana prompt guide for that work, and the logo-specific guide when the deliverable is a mark rather than a scene.

Anything that moves. There is no video path here. Use FLUX 3 Video or MiniMax H3 for motion, and generate your text-heavy stills separately if you need both.

Exact counts and real geography. If the brief specifies twelve bottles on a shelf, or a floor plan that has to match a real building, or a map that a reader will navigate by, this is the wrong tool. The output will look correct and be wrong, which is worse than an obvious failure. Verify counts by hand or use a layout tool.

Long CJK body copy destined for print. Legible is not the same as accurate at paragraph length. For a headline plus a subhead, proofreading is fast. For three paragraphs of Korean or Japanese, typesetting it yourself in a design tool is faster than checking every character.

One more honest limitation: because Alibaba published no weights and no license terms alongside 3.0, commercial usage rights come from the platform you access it through rather than from the model itself. Check the terms of whatever service you use.

Where to Start

If you have never used this model, run the strong jazz poster prompt from the typography section first. It exercises all three type sizes and shows you within one generation how literally the model treats quoted strings. Then swap the quoted text for your own copy and keep the structure.

Qwen Image 3 runs on Morphed with a Generate and Edit toggle on one page on the same page, 12 credits per image either way, output up to 2K, and flexible landscape, portrait, or square sizing. Create an account to run these prompts, or read the AI product photography guide if your next job is commercial stills rather than typography.