SenseNova U1’s infographic fine-tune boosts accuracy 4x on benchmarks

OpenSenseNova has unveiled a specialized fine-tune of their U1-8B-MoT model focused exclusively on infographic generation and understanding. The base model, which already excels in multimodal tasks, underwent an extended multi-task (MT) training phase targeting structured visual output—charts, diagrams, text-heavy infographics. The retooled pipeline prioritizes alignment between visual elements and textual data, a critical capability for enterprise dashboards, automated report generation, and data journalism tools.

The results are striking. On the IGenBench infographic accuracy metric (I-ACC), the fine-tuned model jumps from a baseline of 4.2 to 17.0—a 4x improvement. Chart Understanding climbs from 51.3 to 69.5 (+35%), while Text Rendering accuracy increases from 39.8 to 46.6 (+17%). Overall Aesthetic score remains nearly unchanged at 53.3 (from 53.8), suggesting the visual quality trade-off is minimal. OpenSenseNova has released the model weights and documentation on GitHub, inviting community experimentation and further fine-tuning for niche use cases like scientific figure generation or real-time infographic creation.