SenseTime Open-Sources SenseNova-Vision: Unified Visual Understanding and Generation
On July 13, SenseTime open-sourced SenseNova-Vision (日日新 SenseNova-Vision), a large vision model capable of both understanding and generating images. The model weights, code, and a 50-million-sample visual instruction dataset (SenseNova-Vision Corpus-50M) have all been released.
One Model, Two Capabilities
Most vision models specialize in either understanding (detecting objects, recognizing scenes) or generation (text-to-image synthesis). SenseNova-Vision handles both in a single architecture. According to SenseTime, the model performs well across four core areas: detection, segmentation, depth estimation, and multi-view 3D geometry.
In benchmarks, SenseNova-Vision outperforms specialized models on their own turf. It surpasses semantic-focused models like Youtu-VL on understanding tasks, and beats generation-oriented models like Vision Banana on image synthesis metrics. The key enabler is the model's structured visual understanding and dense geometric prediction capabilities.
50M-Sample Dataset Open-Sourced
Alongside the model weights, SenseTime released the SenseNova-Vision Corpus-50M — a visual instruction dataset containing 50 million samples. This is one of the larger publicly available vision instruction datasets, meaning researchers can fine-tune or build on it without collecting training data from scratch.
SenseTime's Open-Source Strategy
SenseTime built its reputation on computer vision — from facial recognition to autonomous driving. Open-sourcing a unified vision model continues the company's move toward more open releases in recent years. The model can be freely downloaded and used commercially, and the dataset is publicly available as well. For computer vision researchers and developers, it's another solid unified model option to work with.