SenseNova U1 Infographic-V3: A Stunning Open-Source Infographic Generation and Editing Model

SenseTime has officially released SenseNova U1 Infographic Enhanced V3 (SenseNova-U1-8B-MoT-Infographic-V3, or Infographic-V3 for short). It keeps V2’s strong infographic generation ability and — for the first time — adds infographic editing: when the overall image is already good, you no longer have to start over; you simply refine the local text or the overall style.

Model weights: huggingface.co/sensenova/SenseNova-U1-8B-MoT-Infographic-V3

SenseNova U1 Infographic-V3 before-and-after edit example — A before-and-after example of local infographic editing.

Side-by-side original vs edited infographic — Original (left) vs edited (right) infographic.

From “regenerate” to “edit on demand”

In high-information-density scenes, small local fixes — a typo, an icon tweak, a content addition — are almost impossible to avoid. A single text error can make the whole image unusable, forcing users to regenerate and “re-roll” repeatedly. Infographic-V3 raises the bar on delivering usable, structured visual content straight from text-to-image fusion.

Original infographic before local editing — The original infographic prior to editing.

Local text editing

Infographic-V3 supports several editing modes. You can mark a target region for a pinpoint edit, or simply describe the change in natural language. For typos, missing, or duplicated characters it preserves the original layout, illustrations, and visual structure as much as possible.

Infographic restyled to a comic style with modified text — Same infographic restyled to a comic look with rewritten text.

Original infographic before comic-style edit — Original before the comic-style edit.

Original infographic before ocean-themed edit — Original before the ocean-themed edit.

Infographic after removing ocean plastic and adding SAVE OUR OCEAN — After removing ocean plastic and adding a ‘SAVE OUR OCEAN’ slogan.

Localized text edit on an infographic via region annotation — Localized text fix via region annotation.

Infographic with corrected product text via local edit — Product text corrected through a local edit.

Infographic with fixed blurry title text — A blurry title repaired by local editing.

Original infographic before title fix — Original before the title fix.

Infographic with added annual savings target text — An annual savings-target line added locally.

Original infographic before adding savings target — Original before adding the savings target.

Infographic with corrected classical poem text — A classical-poem caption corrected via local editing.

Editing via natural-language prompts

Beyond region annotation, you can specify the exact change directly in a prompt.

Natural-language prompt driven text edit example — Text edit driven by a natural-language prompt.

Infographic edited via natural language prompt — Another natural-language-prompt edit.

Infographic with redesigned slogan text — Slogan text redesigned locally.

Infographic with added curved glass text effect — Curved ‘CURLY & PROUD’ text added onto a glass surface.

Original infographic before curved text edit — Original before the curved-text edit.

Infographic with updated generational AI adoption bar chart — Generational AI-adoptation bar chart updated via editing.

Local content editing

Beyond text, you can also rework local content you are not happy with.

Local content edit example on an infographic — A local content edit example.

Original infographic before local content edit — Original before the local content edit.

Infographic restyled to a LEGO theme — Restyled to a LEGO theme.

Infographic with updated year and headline text — Year and headline text updated.

Global style editing

Infographic-V3 can also adjust the overall style of an infographic while keeping its core information and structure — letting the same content fit different brands, themes, and distribution scenarios.

Infographic in a traditional Chinese style with Song typeface — Migrated to a traditional Chinese style with a Song typeface.

Original infographic before traditional style edit — Original before the traditional-style edit.

Infographic migrated to a light color theme — Migrated to a light color theme.

Global style edit example on an infographic — A global style-edit example.

Original infographic before global style edit — Original before the global style edit.

Infographic with Chinese traditional visual style — Chinese traditional visual style applied.

Original infographic before Chinese style migration — Original before Chinese-style migration.

Infographic in a vintage parchment map style — Vintage parchment-map style applied.

Original infographic before parchment style edit — Original before the parchment-style edit.

Infographic with light-theme migration — Light-theme migration example.

Original infographic before light-theme edit — Original before the light-theme edit.

Global layout editing

It can reorganize content hierarchy and layout while preserving the core message, making the information clearer and adaptable to different sizes and use cases.

Global layout edit example on an infographic — A global layout-edit example.

Generation ability is not sacrificed

Crucially, adding editing did not come at the cost of generation. Infographic-V3 retains V2’s strengths in text-image fusion, complex layouts, and structured visual expression, and still handles high information density, intricate layouts, and diverse visual styles.

WeEdit benchmark score for Infographic-V3 — WeEdit benchmark: Infographic-V3 scores 5.89, best among open-source models.

Qwen-Image-Bench results for the Infographic series — Qwen-Image-Bench: steady growth across the Infographic series.

Trained from the mid-training stage for stable editing

To get stable editing, the team did not only tweak post-training; they went back to the MT (mid-training) stage and redesigned the training path. Around real editing needs they synthesized diverse infographic-editing data and jointly trained text-to-image and image-editing tasks in a fixed ratio. The model then proceeds through MT, SFT (supervised fine-tuning), and RL (reinforcement learning). Experiments show this joint training lets Infographic-V3 keep strong generation while gaining text and style editing — a step from “generate in one shot” toward “visually editable works.”

New editing benchmarks

For the new editing ability the team added WeEdit, a benchmark dedicated to text editing, and adopted the recently open-sourced Qwen-Image-Bench for overall text-to-image evaluation. Infographic-V3 performs well on both, validating its text-modification, local-editing, and comprehensive image-editing skills, with steady gains in overall text-to-image ability. On WeEdit it scores 5.89 on average — the best among open-source models.

Next step: self-check, self-correct, iterate

Infographic-V3 already supports single-turn generation and single-turn editing for an integrated create-then-refine experience — but that is only the start. The SenseNova-U1 architecture natively supports interleaved text-image thinking, so the model may eventually check content and layout after generation, autonomously call its editing ability to fix problems, and optimize through multiple rounds. When generation, checking, and editing form a closed loop, infographic creation moves from “users re-rolling endlessly” toward “the model polishing the finished piece.”

Now open source

SenseNova U1 Infographic Enhanced V3 is now open-sourced globally. Creators and developers are welcome to download and try it, and to keep the feedback coming.

Model weights: Hugging Face
Docs: GitHub
ModelScope: ModelScope

SenseNova U1 Infographic-V3: A Stunning Open-Source Infographic Generation and Editing Model

SenseNova U1 Infographic-V3: A Stunning Open-Source Infographic Generation and Editing Model