OpenSenseNova has released the full version of SenseNova U1.5-Lite, an image generation model that consolidates specialized experts into a single unified architecture. The release introduces task-oriented reinforcement learning and leverages Open-Parameter Distillation (OPD) to merge models trained for text rendering, infographics, aesthetic quality, and editing.

  • Benchmarks show the full version outperforming the preview on Qwen-Image-Bench (60.18), ImgEdit (4.59), and GEdit-Bench-EN (8.26).
  • The model supports native 4K generation with stable global structure and precise local edits.
  • Post-training includes objectives for instruction adherence, visual quality, preference alignment, and edit fidelity.
  • Representations from understanding tasks transfer to generation, improving handling of complex constraints and spatial relationships.

This approach allows users to access specialized capabilities without manual expert selection or routing mechanisms at inference time.