SenseNova U1.5-Lite release: Expert training with OPD distillation

SandyL925 · reddit · 2026-08-21

SenseTime released SenseNova U1.5-Lite, adopting a "specialized in training, unified in delivery" approach. It trains task-specific expert models for text rendering and image editing, then distills them into a single model using OPD (Online Pairwise Distillation), eliminating the need for routing or switching during inference.

Key Improvements:

Benchmarks: Achieved 60.18 on Qwen-Image-Bench, 4.59 on ImgEdit, and 8.26 on GEdit-Bench-EN. The model is open-sourced on GitHub and Hugging Face.

Related event: SenseTime Open-Sources SenseNova U1.5 Lite: 8B Unified Multimodal Model with Native 4K Generation and Editing(14 posts)→

Original post →

More from Multimodal

Multimodal channel →