Skip to main content

SenseTime Open-Sources Unified Vision Model That Handles Four Tasks at Once

SenseTime has officially released and fully open-sourced SenseNova-Vision, a unified visual large model that handles both understanding and generation tasks. This marks a major upgrade in the company's SenseNova large model system, and it's not just another incremental improvement.

What makes it different?

Previous "unified vision" models in the industry often just packaged together multiple expert models—one for detection, another for segmentation, a third for depth prediction. That's not truly unified; it's more like a patchwork. SenseNova-Vision's core innovation is making vision a native capability of a general foundation model, thoroughly integrated into the system. Think of it as teaching a single brain to see, understand, and create, rather than having a committee of specialists.

Image

Outperforming four key domains with one model

In various evaluations, SenseNova-Vision leads extensively across four core visual domains as a single model, matching or even surpassing specialized expert models in each area.

For structured visual understanding—things like object detection, referential detection, OCR, and keypoint localization—it significantly outperforms similar general models. Its performance in complex scenarios like dense small-object detection and long-tail category recognition is particularly impressive.

In dense geometric prediction, the accuracy of depth estimation and surface normal estimation reaches the level of dedicated geometric models, maintaining high stability in both indoor and outdoor environments.

Image

Segmentation capabilities cover general segmentation, reasoning segmentation, and interactive segmentation. Thanks to its strong multimodal understanding, it performs remarkably well in reasoning segmentation and dialogue-based segmentation. With just a single model, it can complete high-quality multi-view point cloud reconstruction and camera pose estimation, achieving leading performance among general visual approaches.

Completely outperforming Vision Banana, with data open-sourced

Compared to semantic-oriented models, SenseNova-Vision achieves comprehensive leadership in visual tasks with high requirements on details such as detection, segmentation, and depth. Compared to the generative-oriented model Vision Banana, it demonstrates a comprehensive generational advantage—surpassing and leading in most metrics in authoritative evaluations. Vision Banana can only address two of the four core areas, while SenseNova-Vision covers all tasks including structured understanding, dense geometry, panoramic segmentation, and multi-view 3D.

Key Points

  • SenseNova-Vision is a unified model that handles detection, segmentation, depth estimation, and 3D reconstruction.
  • It outperforms specialized models in each domain, as well as generative models like Vision Banana.
  • The model and associated data have been fully open-sourced.