SenseNova-U1 Infographic V2: the scope of its 50.3 infographic score
Typography, layout and semantics are evaluated separately, with MoT routing and the BizGenEval hard result kept in scope.
SPECIALIST MODEL / INFOGRAPHICS
An infographic combines semantic planning, hierarchy, icon-to-copy alignment and legible glyph rendering. SenseNova-U1-8B-MoT-Infographic-V2 releases full weights and reports 50.3 on the BizGenEval hard split, a result about specialist infographic generation rather than photography or free-form illustration.
Mixture of Transformers
Routing assigns different parameter subspaces to different visual-language patterns. If p_i=softmax(g(h))_i, a layer can be written as:
y=Σ_i p_i·T_i(h)
Specialist training captures the statistics of headings, data blocks, connectors and decoration. It can still misspell rare names, units or long clauses, so generated graphics remain visual candidates rather than factual sources.
Reading 50.3 in scope
The model card reports 50.3 for V2, 8.2 for Z-Image and 6.3 for Qwen-Image-2512 on BizGenEval hard. The comparison is valid inside that publisher table. Prompt templates, judge and difficult-case composition all shape the result; it does not establish a universal image-quality ranking.
Post-generation review
Every required proper noun, number and unit should be listed before generation and checked against OCR afterwards. Correct characters are only the first gate: title hierarchy, arrow targets, legend colours, reading order and source facts also need review.
What changes create a new result
The 50.3 value belongs to the named BizGenEval hard setup. Changing canvas, prompt template, sampling schedule or judge produces a separate result that cannot inherit the published score.
Infographics became a separate branch of image generation
Early text-to-image benchmarks rewarded semantic similarity and overall aesthetics. Infographics also require titles, numbers, icons and connectors to agree. Infographic V2 is therefore not a generic model with a style word; it is a specialist checkpoint trained further around layout and text-dense composition.
Its 50.3 on the BizGenEval hard split makes that specialization visible. The same official table reports 8.2 for Z-Image and 6.3 for Qwen-Image-2512. The gap is evidence for this publisher benchmark, not proof that V2 also leads photography, portraiture or free-form illustration.
MoT routing for infographic computation
h′=Σ_{e∈TopK(g(h))} p_e(h)·E_e(h)g(h)router scoresTopKexperts activated per tokenE_eexpert-specific transform
Infographics combine imagery, characters, icons and layout relations. MoT routing sends each token to a small expert subset, increasing total capacity without executing every expert for every token.
The 50.3 result belongs to the official BizGenEval hard table. Publication still requires OCR for names, numbers and units plus human review of arrows, legends and reading order.
Paper, code and weights can be followed through SenseNova-U1 source / Infographic V2 weights and benchmark.
The valid scope of 50.3
All three rows come from the same official hard-split table and can be compared there. They do not form a universal image-quality leaderboard.
| Model | BizGenEval hard | Role on this page |
|---|---|---|
| SenseNova-U1 Infographic V2 | 50.3 | specialist infographic model |
| Z-Image | 8.2 | same-table general baseline |
| Qwen-Image 2512 | 6.3 | same-table general baseline |
Checks that start after the image is finished
List every proper noun, number and unit from the prompt, then run OCR on the result. Correct characters are only the first gate: title hierarchy, arrow targets and legend colors must also remain consistent.
Long multilingual strings and mixed units are particularly vulnerable to bad line breaks and substitutions. Keep editable source data; treat the generated PNG as a layout candidate rather than an unchecked factual artifact.
A high score still misspells text
A grader is not a reader. Rare names, long units, footnotes and dense tables can still contain pseudo-characters, substitutions and bad wrapping.
The 50.3 result belongs to the BizGenEval hard setting listed on the model card. Changing canvas, prompt template, sampling schedule or judge creates a separate result and cannot inherit the published score.
Papers and public files for SenseNova-U1 source
Model names, numbers and limitations trace to the papers, repositories or model cards below.