SenseNova U1.5-Lite Officially Released

SenseNova has officially released U1.5-Lite, an image generation and editing model. This model adopts an approach where expert models specialized for specific tasks such as text rendering, infographics, aesthetic quality, and image editing are trained first, then integrated into a single model via the OPD (On-Policy Distillation) technique. This enables service delivery with a single model during inference without routers or expert switching, aiming for 'specialization during training, integration during deployment'.

Key improvements are as follows:

• Complex instruction adherence: Enhanced ability to simultaneously handle multiple constraints such as subject, count, spatial relationships, text, and layout.
• Text rendering and 4K generation: Improved Chinese/English text rendering in posters and infographics, with support for native 4K generation featuring stable global structure.
• Editing precision: Strengthened capabilities in preserving subject identity, geometry, and unedited regions, enabling multi-reference editing and precise local edits.

In benchmark results, it recorded 60.18 points on Qwen-Image-Bench (with PE applied), demonstrating performance improvements over previous versions. Additionally, it possesses structural strengths where representations learned in visual understanding transfer to generation and editing tasks. JSON-structured supervision is used as a training mechanism for controllability, while the user interface remains natural language-based.