SenseTime shipped SenseNova U1 Pro

At the 2026 World Artificial Intelligence Conference on July 18, SenseTime introduced SenseNova U1 Pro, the flagship model in its SenseNova U family.

What it can actually produce

Every piece on the product page comes with the Chinese prompt that generated it. The most telling one is this, because it describes no layout at all.

帮我做一个周六去周日回的溧阳旅游攻略，涵盖景点、住宿和小吃

The output from that one conversational request — roughly "make me a Saturday-to-Sunday Liyang travel guide covering sights, lodging, and street food." The zones, the icons, the itinerary timeline, the hierarchy of the food and lodging blocks: the model decided all of it.

The other category is long prompts that specify the layout down to the millimeter. This one is fully visible, and the names of all five modules come straight from the prompt:

生成一幅以手提包为核心主题的信息图，标题为《皮囊解析：被携带的自我》，探讨个人物品与自我身份的关系，包含甄选法则、容器、内衬秘密、生活沉淀、磨损情结五个主题模块

All five specified modules are there, and both the title hierarchy and the body-copy fine print read clearly. One instruction got dropped, though: the prompt asked for a handbag and the model drew a backpack — the specimen card inside the image even labels it "canvas backpack."

Those intimidating long prompts probably weren't written by hand

The truncated long prompts on the product page share a trait: they push the entire design brief into natural language. Aspect ratio, primary and accent colors, type family, column count, how much white space, which element dominates the visual hierarchy and which comes second, and a requirement that all text be in Chinese. The prompt behind "Silver Halide: How Light Becomes a Photograph" runs about 1,800 characters and includes exclusions like "no color cards, complex background textures, or photographic elements."

Nobody writes like that off the cuff. The open-source docs supply the missing step: before generating an infographic, have the model expand your plain-language request into a full visual spec, then feed that expansion into image generation. SenseTime published a separate best-practices document just for this prompt-enhancement step.

Once you know that step exists, the long prompts on the product page make sense: they're intermediate artifacts, not what a user typed.

Not everything is an infographic

Two other sections of the product page hold design and scene work, judged on whether it looks good rather than how much information it packs in.

Lixia, the Start of Summer, from the 24 solar terms. A watermelon kite, the word SUMMER pulled apart, four thin strings dangling facts about the season, and the date and lunar calendar laid across a grassy slope.

A Quanzhou city poster in woodcut style — two-color line work in black and orange-red, with a grainy noise texture.

A Dunhuang Museum heritage exhibition poster. Mural bodhisattvas fill the background; the gold calligraphic title runs vertically down the right side in staggered, overlapping columns.

A retro cola ad. An oversized yellow-on-black headline eats half the layout, with three planes of depth: figure, ice bucket, and bottle in the foreground.

An energy drink hero shot — the can suspended in backlit dust, brush lettering aligned to the can's own markings. This is commercial campaign quality.

A photorealistic portrait; the prompt asked for visible pores. Light through the leaves dapples her cheek and collarbone, loose hair is caught by wind, and there's dust hanging in the air.

An architectural rendering: a flowing shell-form timber roof over floor-to-ceiling glass. The icon wayfinding strip along the bottom and the carved stone signage on the right are part of the generated image, and the serif headline sits against the sky.

Not a single word in the frame. A floating birdcage holds an entire miniature spring valley; the door is open and the stream spills out as a translucent ribbon of water into the real landscape below.

The dense text-and-image side is a completely different test: hold the layout together and keep the type correct at high density.

A vertical data visualization with six numbered modules, separated by hairlines, most of them carrying their own chart or icon.

A landscape academic review diagram on the microscopic mechanisms and multi-scale closed-loop lifecycle of all-solid-state lithium batteries, built for a research presentation.

Product assembly instructions in blueprint style — white line work on deep blue, with dimension lines, structural lines, and scale marks all in white.

A children's science illustration. The prompt specifically ruled out columns and info cards, asking instead for one continuous story scene with the facts woven into the environment.

How it keeps the text from smearing

To see why SenseTime is willing to put "very low text rendering error rate" on a capability list, you first have to look at how images got built the old way.

For years, the standard multimodal architecture split the job between two components. A vision encoder did the seeing, turning images into features the model could read. A variational autoencoder did the painting, expanding a short string of numbers the model produced back into pixels. An adapter bolted the two sides together.

There are two seams: one between understanding and the intermediate representation, another between that representation and generation. The two ends were built on different representations to begin with, so every crossing requires a translation — and the finest strokes in the image are the first thing sacrificed.

The architecture is a joint effort between SenseTime and Nanyang Technological University. A preview went public as a blog post back in March, but the model weights weren't actually open-sourced until April 27.

If you delete the component that paints pixels, who paints?

This is the obvious question, and the answer is a training approach called flow matching, applied directly at the pixel level.

So the training objective splits into two parallel tracks: autoregressive cross-entropy for the text side, pixel-level flow matching for the visual side. That's what's actually happening inside the "generation" box on the right of the architecture diagram.

Does image quality suffer for it?

SenseTime's answer is a set of image reconstruction scores — feed in an original image, reconstruct it, and measure how close the result is.

Efficiency is the other half of the case: at equal training data volume, NEO-unify reached higher performance than the Bagel baseline using fewer training tokens. That's the other reason SenseTime is willing to bet on this route.

The weaknesses the open-source release admits to

The open-source repo includes a "work in progress" section listing four known limitations. One of them lands squarely on the same battlefield as that "very low error rate" claim.

What's open-sourced is the Lite line; U1 Pro is the flagship, at a different scale, so the conclusions may differ. But U1 Pro has published no benchmark scores at all, and the closest available results are the open-source version's numbers on infographic benchmarks like BizGenEval and IGenBench.

It can change one thing without redrawing everything

The architecture solves whether text and image can align. The other half of the problem: what happens when the first image isn't right.

The answer lies in what the model writes down before it starts drawing. The open-source inference examples show the full process. Give it a photo of tea steeping in a glass and ask it to draw what this looks like an hour later, and it outputs six reasoning steps before it touches a pixel.

Step five is the mechanism behind changing one thing: before it acts, the model writes itself a do-not-touch list. With that list in hand, editing shifts from rerolling the whole image to a constrained local modification. The open-source repo has three more examples with the same structure — a beanbag chair after long use, an egg moved from fresh water to heavy brine, a green banana after ripening.

Close those six steps into a cycle and you have what SenseTime calls the long-horizon generation loop. Around one complex goal, it can run dozens of rounds this way.

The third version of the open-source infographic model, released July 16, targets editing specifically: local text, local content, global style, and global layout, with an explicit claim of precise text repair inside dense blocks of copy. That's exactly the "the date line smeared, fix only that line" capability.

How far the loop stretches

Blade of the Sand Shadow, shown at the launch, is the answer. It's a demo project covering an entire film pre-production package.

The model first produced the full visual bible on its own: world-building, hero and antagonist character designs, three-view turnarounds and expression sheets for the monster, environment designs, and concept art. On top of that foundation it then output 22 consecutive storyboard frames in one pass, each annotated with camera direction. Faces, costumes, and lighting have to stay consistent across all 22 frames — something random assembly cannot do. It requires a locked set of designs that every subsequent frame is bound to. Paired with SenseTime's video tools, the result is controllable delivery of character, plot, and visual language.

Far left is the world-building text block, followed by character designs (traveler form and combat form), monster turnarounds and expression sheets, environment designs, and concept art. The bottom two rows hold all 22 storyboard frames, each with camera notes. The source is a thumbnail overview from the launch materials — the detail needs the full-size version.

Two more pieces from the WAIC stage

1. A 4:1 ink-wash scroll

A World Convened in Intelligence, made for WAIC's ninth anniversary, is rice-paper ink-wash style at an ultra-wide 4:1 ratio. From a distance it reads as one continuous piece of Eastern landscape narrative, rivers and mountains flowing into each other. Zoom in and every conference from 2018 through 2026 has its own landmarks, events, and fine-print annotations holding the picture up. You can pick out the lighthouse and AI PARK from "2018, the first WAIC launches in Shanghai," plus "2025: One World, One Shared Future," "2026: Intelligent Partners, Creating the Future Together," and a run of names including the Global AI Governance Action Plan, "100,000 m² of exhibition space," "1,100+ companies," and the SAIL, Yunfan, and Youth Paper awards.

Top is the full scroll; below are two detail crops — the 2018 inaugural venue on the left, the 2025–2026 stretch on the right. The test for an image like this is whether it reads as one continuous piece from afar and every character still holds up under magnification. This is the thumbnail from the launch materials; the scroll's real detail needs the full-size original.

2. A pre-match prediction report for the World Cup final

Published July 18, before the 2026 World Cup final had kicked off. The image is a pre-match analysis and score prediction: Spain versus Argentina, with a predicted scoreline of 1–2. Each side's route to the final, key player matchups, tactical formations, passing networks, and touch heat maps are all compressed into a single image — the full run from information gathering through analysis and reasoning to complex visual delivery.

There's a related track record here, though it belongs to a different product: SenseTime's office assistant Raccoon predicted a month before kickoff that Cape Verde would be the tournament's biggest dark horse and advance from the group, which SenseTime says placed it second among 12 models making predictions. Raccoon made the prediction; U1 Pro turned the result into the image.

PREDICTION and the 1–2 scoreline sit dead center, flanked symmetrically by passing network diagrams, touch heat maps, goal analysis charts, knockout routes, player cards, and two tactical formation diagrams. The source is a thumbnail overview from the launch materials.

The open version is downloadable now; the full release comes in August

You can't get U1 Pro yet, but the open-source SenseNova U1 Lite line, built on the same architecture, has been out since April under Apache 2.0 — commercial use allowed.

Two-plus months of open-source iteration have concentrated on infographics and editing.

On usage, SenseTime's figures: in June, open-source users generated nearly 3× as many images per person per day as in May, and as of mid-July, the model plus SenseNova-Skills — a repo packaging it into ready-to-use skills — had passed 8,000 total GitHub stars.

An image-understanding model shipped too

SenseTime also announced SenseNova-Vision at WAIC, built on the same native unified architecture. It turns classic vision tasks like instance segmentation and object detection into native capabilities of a general-purpose large model, rather than something stitched together from multiple specialist models.

Four demos: ultra-dense object segmentation separating a school of overlapping fish, seeing through mirror reflections to recover spatial geometry, zero-shot comprehension of a game it has never seen, and judging real spatial structure in optical illusion images.

When you can actually use it

The U1 Pro preview is now in invite-only testing, and the invite pool will widen as inference capacity grows. The public release is planned for August, with pricing and API service launching alongside it. In the meantime, the same capabilities are already running inside SenseTime's office assistant Raccoon and its video creation tool Seko.

At the World Artificial Intelligence Conference on July 18, SenseTime launched SenseNova U1 Pro. It puts seeing, generating, and editing inside one model, looping dozens of rounds around a single goal until it produces something usable as-is. Output runs up to 8K, so small type still holds at print size.

Getting there took a new architecture underneath. To see what changed, you first need to know how images used to get built.

XiaoHu wants a marathon poster: headline, the date "October 26," meeting point, QR code bottom right — all four locked. The diagram below runs that same request through two different architectures.

Put "change one thing" into concrete terms: one poster, three rounds of client notes.

All of the above are SenseTime's own figures. U1 Pro has published no benchmark scores and no third party has tested it. The closest available results come from the open-source version on the same architecture: image reconstruction scores of 31.56 / 0.85, slightly below the Flux VAE baseline (32.65 / 0.91) — dropping that compression component does cost reconstruction fidelity. The open-source release also states plainly that rendering still breaks in text-dense scenes.