Prompt cap raised to 4.5k tokens — a whole complex layout, spelled out in one go

The middle bar is the number spelled out in Alibaba Cloud Bailian's API docs: the qwen-image-2.0 line caps prompts at 1300 tokens, other models at 800, with anything past that simply truncated.

When prompts were short, a complex layout had to be drawn in pieces and stitched together. Nine subjects, each with its own title, body text, formulas, and character actions — a thousand characters couldn't even cover half of that. Raising the cap to around three thousand is what turns this kind of job from piecemeal collage into a single pass.

Qwen's proof point here is a nine-panel grid, whose description ran to 3.7k tokens.

All nine panels were generated from a single 3.7k-token prompt, with no stitching in between. The subjects: a tunnel safety-distance comic, spatial geometry, a stylistic analysis of the classical essay "Chu Shi Biao," projectile motion, parasitology, a medical diagram, group theory's Sylow theorems, a bank internal-controls infographic, and a cell/DNA structure comparison. Source: Qwen launch page

Pull out any single panel and it stands on its own. The upper-middle one is a complete spatial-geometry teaching page — problem, premises, theorem, and conclusion, all present.

The information density of a single panel matches that of a full teaching slide. Source: Qwen launch page

Worth noting: the derivations in these panels aren't just for show. The group-theory panel works out how many Sylow 3-subgroups a group of order 72 has — factoring 72 into 8 times 9, applying the constraint that the count must be congruent to 1 mod 3 and divide 8, and landing on an answer of 1 or 4. The whole chain holds up. The physics panel derives the angular velocity formula from an object's initial velocity at the point of release, and the landing-time and displacement substitutions in between check out too.

Spatial control, two directions: side-by-side layout, and layer-by-layer nesting

The nine-panel grid in the last section proved lateral spread. The depth example is a single prompt producing four nested layers: a VSCode editor on the outside, a Qwen chat window open inside it, a WeChat screenshot inside that chat, and a pour-over coffee process poster shared inside that WeChat conversation.

Four nested layers, and each layer's interface elements hold up on their own: VSCode has its file tree and status bar, WeChat has its contact name and timestamps, the poster has its four numbered steps. Source: Qwen launch page

These two directions serve different jobs. Lateral spread handles multi-column layouts — newspapers, exam sheets, teaching wall charts, product comparison pages. Depth handles product-demo decks and interface sequences, showing how a feature jumps from one piece of software to another. What used to take three screenshots stitched together now takes one prompt.

Smallest legible text hits 10px — enough to lay out a full newspaper page or a whole paper

Three kinds of work depend most on this capability.

1. Formula-dense academic layouts

The hard part of LaTeX layout is that superscripts, subscripts, braces, fraction bars, Greek letters, and multi-line alignment all show up at once — miss any one position by half a grid step and it shows.

A full two-column page where formula sub/superscript placement, multi-line alignment, and special symbols all follow LaTeX conventions. The header reads "Fake PDF" — this page was generated to spec for a fictional paper. Source: Qwen launch page

2. Newspaper layout

A newspaper front page has a lot that has to work at once: the calligraphic strokes of the masthead, the dividing lines in the date bar, the headline hierarchy across several stories, a chart card in the middle, a reading-guide column on the right, and the baseline alignment of body text in every column. Miss any one and it stops looking like a newspaper.

The whole page's content is fictional, per the prompt. Beyond the dense small text, the paper's creases, shadows, and print texture were generated too. Source: Qwen launch page

3. High-density science infographics

This kind of image crams body text, leader-line labels, a legend, and a scale bar all together, with text pushed down to the smallest size — and every label still has to point to the right spot.

Eight numbered sections, plus a world map of migration routes and a size-comparison scale, all with text at the smallest size tier. Source: Qwen launch page

Pores, hair strands, and material textures all hold up to a close look

Portraits: the harder the light, the harder to hide flaws

Hard light pushes every bump and pore on the skin into plain view, and shadow edges can't be allowed to blur. The image below also has a floral-branch shadow cast across the face — the detail right at the light-dark boundary is where the real test is.

A petal-shaped shadow falls across the face, while a row of ear cartilage piercings, the highlight on the lip gloss, and the print on a floral shirt all stay crisp at the same time. Source: Qwen launch page

The natural-light shot tests something else: hair strands. Wind-blown wisps have to separate strand by strand against backlight, with light visibly passing through them.

The Swiss cheese plant and ferns in the background stay softly blurred by depth of field, while the loose strands of hair in the foreground separate individually. Source: Qwen launch page

Materials: four completely different surfaces

The non-portrait examples show the range better. For embroidery to work, three things have to land at once: feather layering has to come through in stitch direction, the thread needs the actual sheen of silk, and the linen weave underneath has to show through.

Thread direction, color gradation, and the base fabric's weave are three distinct textures, and each one holds up on its own within the same image. Source: Qwen launch page

Animal fur runs on a different logic: black fur has to keep its layering in the shadows, and white fur can't blow out under a flash.

A nighttime flash shot where the black coat's layering and the white coat's downy texture both hold up at once, with streetlight bokeh in the background. Source: Qwen launch page

Rock art tests the relationship between pigment and stone: mineral color has to seep into the pores of sandstone, and it needs the mottling of weathering and flaking.

Ochre-red pigment painted on sandstone, with the rock's cracks, grain, and pigment flaking all rendered together. Source: Qwen launch page

This last one is an extreme macro shot: the thick stacked strokes of oil paint, every individual bristle on the brush, and the scratches and reflections on its metal ferrule.

The paint's thick-thin buildup, the separated strands of hog-bristle brush hair, and the canvas weave peeking through at the edge — three layers of texture, all near the same focal plane. Source: Qwen launch page

One model both draws from scratch and edits the image you already have

Adding handwritten annotations to an existing book page

The hard part is how real the handwriting looks. Red-pen strokes need pressure variation, circled words need the irregular hand-drawn oval of an actual pen stroke, arrow tails need a slight hook back at the end, and it all has to sit on top of the printed text rather than float off to the side.

Before: a book page about Xu Xiake, with a travel-route map on the lower half

After: more than a dozen red-pen annotations, underlines, circles, and arrows layered on top

Restoring missing sections of a damaged painting

The requirement here is that the brushwork match the original after restoration, with no sense of a patch job.

Before: white gaps of loss across the whole surface, horizontal creases, brown stains

After: even silk tone throughout, with the inscription seal and composition's brushwork held in their original places

Overlaying a professional annotation layer on a real photo

The input is an ordinary macro photo, and the output has to be usable directly as an academic figure: the subject can't distort, and labels have to point to the right spots.

Before: two damselflies on a blade of grass

After: taxonomic information, morphological labels, zoomed-in detail, and a scale bar added

Give it a character and a scene, and it lays out a full storyboard

Qwen's example set here is four images. The first three are source material: a character reference, a scene, and one key shot.

Character reference: black-frame glasses, a beige suit, a wine-red turtleneck. Source: Qwen launch page

Scene: an old mountain town at dusk, with black tile roofs, timber construction, stone-paved streets, and layered distant mountains. Source: Qwen launch page

Key shot: black and white, a backlit silhouette with arms spread wide in a curtain of rain. Source: Qwen launch page

The fourth image strings the first three together into a complete storyboard.

Twelve panels, each labeled with a shot number, duration (ranging from 3s to 1s), shot type, and action description — wide shot entering the town, medium shot walking, close-up reading a note, over-the-shoulder shot spotting a dark figure ahead, handheld turn, insert shot of a hand pushing a door open — with the camera-movement notation (hold, move, push in, follow) marked in the bottom-right corner. Source: Qwen launch page

This image draws on every capability covered so far: the same character carries through all 12 panels, the scene and weather advance with the narrative (entering the town, rain starting, the door pushed open), each panel's black label bar packs in dense small text, and the whole thing is a lateral-spread layout in its own right. Panel 10 is the rain silhouette shown above.

The old workflow for this went: generate the character reference, then the scene, then each panel one at a time, then pull everything into design software to lay out and annotate. Now the whole set comes out in one pass.

12 languages, 100-plus styles, mainstream software interfaces — and it pulls in today's information online

Multiple languages, each paired with its own local layout conventions

This dimension covers native rendering in 12 languages and more than 100 art styles. The three examples here each take a completely different route. The Japanese one is a manga title page — black-and-white screentone, vertical dialogue boxes, and sound effects all have to line up together as one visual language.

Vertical Japanese text, manga paneling, and screentone style all hold up at once. Source: Qwen launch page

The Korean example is an e-commerce product page: spec icons on the left, a model in the middle, a parameter table on the right, four lifestyle shots along the bottom — a complete product-page layout.

A dress product page: six line-drawing icons with labels, five color-swatch thumbnails, a spec table on the right, and four styled-outfit shots along the bottom. Source: Qwen launch page

The Spanish example switches to handwriting: whiteboard-marker lettering in three colors, plus a neural-network diagram and math formulas.

A handwritten whiteboard explanation of multilayer perceptrons, in five numbered sections: basic structure, forward propagation, loss function, backpropagation, and characteristics — including a neuron-connection diagram and formulas. Source: Qwen launch page

Interface simulation

Web, game, and livestream interfaces each follow their own set of control conventions. This capability feeds directly into product prototypes and demo mockups.

A developer-tool interface: menu bar, project tree, syntax-highlighting colors, and the log format in the console at the bottom. Source: Qwen launch page

A food-delivery app's listing page: phone status bar, search box, filter tags, four restaurant cards (each with a product photo, rating, monthly order count, delivery time, and discount coupon), a coupon banner and five navigation tabs at the bottom. The information density here is even higher than the developer-tool example. Source: Qwen launch page

Pulling today's information online

The model's built-in world knowledge has a cutoff date; going online is what fills in anything newer. Asked for today's Hangzhou weather forecast, it comes back with a complete weather-and-sightseeing guide.

Today's temperature range, weather, wind, and humidity, paired with a temperature curve — the lower half of the page has four attraction cards and a one-day route. Source: Qwen launch page

Recognizing specific IP characters and reworking them

This isn't just about recognizing a face — the model has to carry over the style that character is supposed to be drawn in, too.

Qi Baishi rendered with photorealistic texture, Van Gogh kept in thick oil-paint brushwork — two completely different rendering styles placed into the same livestream lighting setup. Source: Qwen launch page

Where to use it now, what it costs, and whether it's open source

Qwen's launch page never mentions any of these three things. This site checked them directly.

In Alibaba Cloud Bailian's pricing table for Qwen text-to-image models, qwen-image-3.0-pro already sits in the top row, marked "free for a limited time" in both the mainland China and international regions. For comparison, the previous-gen qwen-image-2.0-pro cost ¥0.5 per image.

Bailian's API documentation hadn't been updated as of that day — the newest entry in the model list was still the June 22 version of qwen-image-2.0-pro. The pricing table is already live; the docs will need to catch up. On the web, it's available through Qwen Chat by selecting "image generation."

Open source is a separate story. In the Qwen organization's image-model repositories on Hugging Face, the last weight release was Qwen-Image-2512, dated December 30, 2025. February's Qwen-Image-2.0 never got a weights repository — the corresponding entry in the GitHub changelog only links to the blog post and Qwen Chat. 3.0 currently has neither.

When using it, write longer prompts. This generation raised the cap to 4.5k tokens precisely so you can spell out every word that needs to appear and where every block needs to sit.

This generation raised the single-prompt cap to 4.5k tokens, roughly 3,000 Chinese characters. For the exact same job, the level of detail you can spell out is on a whole different scale.

Once the prompt can hold it, the jobs it can take on change too:

• Text down to 10px holds up, so a full newspaper page, a whole LaTeX-formula paper, or a dense field guide can all be laid out
• Give it a character reference plus a scene and it lays out a 12-panel storyboard directly, each panel labeled with shot type, duration, and camera movement
• Editing and drawing live in the same model: layer red-pen annotations onto a book page, patch the gaps in a damaged classical painting, add an academic annotation layer to a real photo
• 12 languages laid out in their own local conventions, mainstream software interfaces simulated, and even today's weather pulled in online

Now that prompts can run long, the next question is: with that many things packed into three thousand characters, how does it arrange them all on one canvas?

Say you've got two jobs on your desk: a nine-panel teaching grid, and a product-demo image set. Each one draws on a different one of the model's spatial-control capabilities — one spreads sideways, the other nests inward.

Lateral spread decides how many things one image can hold at once; depth decides how many layers of interface it can nest inward. Newspapers, exam sheets, and teaching wall charts lean on the former; product-demo decks and interface sequences lean on the latter. Diagram by this site.

This generation only really has one number to show for itself: how much you can say in one go. The longer the bar, the more it can pack into a single image.