SenseNova-Vision 7B unifies detection, OCR, and depth in one open model
SenseTime released SenseNova-Vision, a 7-billion-parameter open-weight multimodal model that handles object detection, OCR, segmentation, depth estimation, and pose through natural-language prompts—all without task-specific output heads.
A single vision model that takes plain-English instructions and returns detection boxes, depth maps, or OCR results without swapping out task-specific modules—that's the pitch behind SenseNova-Vision, a 7-billion-parameter multimodal generative model released this week by SenseTime.
SenseNova-Vision is a mixture-of-tasks (MoT) architecture that unifies object detection, optical character recognition, instance segmentation, keypoint extraction, surface-normal prediction, depth estimation, and camera-pose estimation under a single decoder. Instead of maintaining separate output heads for each task, the model interprets natural-language prompts and generates text, images, or hybrid outputs depending on the instruction. A user can ask for bounding boxes around objects, request a depth map, or pull text from a scanned receipt—all through the same interface.
The model ships with open weights on HuggingFace under Apache 2.0 licensing. SenseTime published a live demo and a GitHub repository with inference code, including example prompts for each supported task: "Detect all people in this image," "Extract text from the sign," "Generate a depth map." The MoT design means practitioners can fine-tune or prompt the model for new vision tasks without retraining task-specific branches, a workflow positioned as more flexible than stacking narrow models.
Because SenseNova-Vision runs locally from open weights, users control what prompts go in and what outputs come out, with no server-side safety layer. The GitHub repository includes a ComfyUI custom-node integration for workflow builders who want to wire detection or depth into image-generation pipelines.
