Connect a camera or video to an AI detection, segmentation, or classification model and watch the results live — all by dragging and linking boxes in a native desktop app. No machine learning experience required.
Quick start · Highlights · How it works · The nodes · Roadmap
# 1. Install dependencies (CPU build — works on any computer)
.\.venv\Scripts\python.exe -m pip install -r requirements.txt
# 2. Start the desktop app
.\.venv\Scripts\python.exe app.pyThe native PySide6 desktop window opens with the node editor. Configure the nodes, click Run,
and the detections appear in a live preview window embedded in the app.
Press Stop, or close the preview window (q / Esc) to stop.
The editor is now a native desktop app — there is no browser or local web server involved. Your graph is saved to
workflow.jsonnext to the app and reloaded on the next launch. The legacy browser editor (python main.py, served athttp://127.0.0.1:8000) still works if you prefer it.
Have an NVIDIA GPU? Run
.\scripts\install-gpu.ps1instead of step 1 for a big speed-up. See GPU acceleration below.
The running version is shown in the top bar next to VisoNode.
The version source of truth is web/version.json. Update that file before a release. Python also reads it for:
.\.venv\Scripts\python.exe main.py --versionWhen the backend is restarted, /api/app/version also reports the app version and current Git commit.
- No code. Build the whole pipeline by connecting nodes in the browser.
- 100% local. Your camera and video never leave your computer — nothing is sent to the cloud.
- State-of-the-art detection, segmentation, and classification. Powered by Ultralytics YOLO26, Roboflow RF-DETR, and Meta's official SAM 3 concept segmentation.
- CPU or GPU. Works on any machine; uses your NVIDIA GPU automatically when available.
- Live feedback. See frame rate, object count, and the active device in real time.
- Flexible inputs. Webcams, stream URLs, capture devices, or local image/video files.
A workflow is just a chain of nodes:
Input → Object Detection, Object Segmentation, or Object Classification → Class Filter → OpenCV Preview → Alert Output
Each node does one job: read frames, find objects, segment object masks, classify the frame, keep only the classes you care about, draw the results, and log alerts. The results appear in a separate window, with boxes, labels, classifications, and masks drawn over the video:
VisoNode runs entirely on your own computer:
- The browser is just the editor. You use it to lay out the nodes and change their settings.
- Python does the real work behind the scenes: opening the camera or file, running the selected AI model, filtering results, drawing boxes, and logging alerts.
When you click Run, the browser hands your workflow to Python, and Python opens a native preview window showing the live detections. In the editor, each node turns green and shows RUNNING, and the top bar reports the live frame rate, object count, and active device.
First run note: the first time you use Object Detection, it automatically downloads the YOLO26 model weights (for example
yolo26n.pt). This happens once and may take a moment. RF-DETR also downloads its selected starter weights the first time you run an RF-DETR model. To use fine-tuned RF-DETR weights, choose RF-DETR and set the optional checkpoint path. SAM 3 uses Meta's official implementation. Request access tofacebook/sam3on Hugging Face, runhf auth login, then choose SAM 3 concept segmentation. You can also set a local checkpoint path in Object Segmentation.Meta's official SAM 3 prerequisites are stricter than YOLO26: Python 3.12+, PyTorch 2.7+, and a CUDA-compatible GPU with CUDA 12.6+.
The Quick start installs the CPU build, which works on any computer but is slower. If you have an NVIDIA GPU, install the GPU-enabled build instead. This script installs a CUDA build of PyTorch, installs the app, and runs a quick test to confirm the GPU is detected:
.\scripts\install-gpu.ps1By default it installs the cu128 driver wheels. To target a different CUDA version:
.\scripts\install-gpu.ps1 -CudaWheel cu126To check whether an existing install can see your GPU:
.\scripts\check-gpu.ps1| Node | What it does |
|---|---|
| Input | Chooses where frames come from. Camera mode takes an OpenCV camera index, a stream URL, or a capture source plus resolution. File mode takes a local image or video path and can loop videos. |
| Object Detection | Runs an Ultralytics YOLO26 or Roboflow RF-DETR detection model. Lets you pick the engine, model size, confidence threshold, and CPU/GPU device. |
| Object Segmentation | Runs an Ultralytics YOLO26-seg, Roboflow RF-DETR-Seg, or Meta SAM 3 segmentation model. SAM 3 takes comma-separated noun phrases such as person, car, bottle; all engines return masks or boxes, labels, and confidence scores downstream. |
| Object Classification | Runs an Ultralytics YOLO26-cls image classification model and passes the top label and confidence score downstream. |
| Class Filter | Keeps only the object types you list — for example person, car, dog. |
| OpenCV Preview | Draws masks, bounding boxes, classification labels, and object labels in a native preview window. |
| Alert Output | Logs detection events on the backend, with a cooldown so you aren't flooded. |
The Object Detection node has a Device setting:
| Setting | What it does |
|---|---|
| Auto | Uses your NVIDIA GPU if one is available, otherwise falls back to CPU. |
| CPU | Always runs on the CPU. Works everywhere, slower. |
| CUDA device | Forces the GPU. Requires a working CUDA + PyTorch install, and reports an error if the GPU can't be used. |
If you're not sure, leave it on Auto.
Planned improvements:
- Save and load workflows from a file.
- Built-in presets for RTSP / IP cameras.
- Use a monitor / screen as an input source.
- Extra inference nodes for ONNX and TensorRT.
- More output nodes: webhooks, database logging, and snapshot saving.
This project is licensed under the GNU Affero General Public License v3.0 — the same open-source license used by Ultralytics. See LICENSE.