Run detection, depth, and segmentation from one vision model

SenseNova-Vision is a single model that handles a wide range of vision tasks through one instruction interface, instead of a different specialist model for each job. You give it an image and a request, and it returns the right kind of output for the task.

The range is the point. It does structured understanding like object detection, keypoints, OCR, and GUI grounding as text with coordinates. It does dense geometry like depth estimation and surface normals as images. It does several kinds of segmentation, including referring and reasoning masks where you describe the thing you want cut out. It even does multi-view geometry like point-map reconstruction and camera-pose estimation. All of that comes from one set of weights, so you learn one tool instead of five.

The code is on GitHub at OpenSenseNova/SenseNova-Vision. Clone it, run the setup script, and pull the weights from Hugging Face. A single example command runs a task end to end, for instance binary segmentation of a person in an image, and there is a Gradio web demo you can launch locally to click through tasks interactively.

Good first tests are the ones a single-purpose tool would need its own model for: ask it to segment an object you describe in words, then have the same model return a depth map for the same photo. Weights are open under a non-commercial license, and because it is a 7B vision model, expect a real GPU rather than a laptop.

It is a 7B model, so local runs want a capable GPU and are not instant. The weights are non-commercial, which is fine for testing and personal projects but blocks commercial use. And a generalist that spans this many tasks will not always beat a dedicated model tuned for one of them, so judge it on the breadth, not on topping any single benchmark.

• Detect and localize objects, keypoints, or text without a separate detection model.
• Cut out an object by describing it, instead of drawing masks by hand.
• Get a depth map or surface normals from a single photo for 3D or relighting work.
• Reconstruct point maps and estimate camera pose across multiple views.
• Prototype a whole vision workflow with one model before wiring up specialists.

Try the demo

View model page