> For clean Markdown of any page, append .md to the page URL.
> For a complete documentation index, see https://docs.ovalbee.com/llms.txt.
> For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://docs.ovalbee.com/_mcp/server.

# Try a pretrained detector

Before labeling thousands of images, it's worth an hour to learn whether a ready-made detector already reacts to your material — and in whose words. The **Run EdgeCrafter Detector on Images** template runs a pretrained detector over a dataset and gives you its boxes to look at.

```mermaid
flowchart LR
  dataset["Dataset"] --> predict["Predict EdgeCrafter"]
```

**Nodes:** [Dataset](/node-library/nodes/dataset-ops-dataset) and a prediction node — [Predict EdgeCrafter](/node-library/nodes/models-edgecrafter-predict-edgecrafter), [Predict YOLO](/node-library/nodes/models-yolo-predict-yolo) or [Predict DEIM](/node-library/nodes/models-deim-predict-deim).

#### Point the flow at your images

Set the Dataset node to your images. They don't need any labels.

#### Choose a model

In the prediction node's **Model** input, pick one of the pretrained models from the catalog. Smaller ones are faster, larger ones are more accurate — each node's page in the Node Library compares them. Optionally tick only the **Predict classes** you care about.

#### Set the confidence threshold

The node's **Config** holds the inference settings, including the confidence below which boxes are dropped. The default is a good start. Lower it to see what the model is unsure about, raise it to keep only confident boxes.

#### Run and look

Run the prediction node. It writes the images with the model's boxes into a new predictions dataset, which the node previews on the canvas. Open the preview to page through the results.

The run needs a GPU worker, and the first run of a model can take a little longer while its image is pulled.

## What to look for

Two questions, answered by paging through the predictions:

* **Does the model react to your objects at all?** If it finds most of them, you may need far fewer labels than you planned — or none, if you only need what it already knows.
* **In whose vocabulary?** Pretrained detectors predict their own classes, such as the 80 COCO classes. If yours are *scratch* and *dent*, the model can't name them, and you'll need to fine-tune. If yours are *car* and *person*, you may only need to map or filter its classes.

**See it running:** [live example](https://app.ovalbee.com/public/flows/run-edgecrafter-detector-on-images) · [case study](https://ovalbee.com/case-studies/detection-on-images-mini-traffic)

## Next steps

* **Pre-label with it.** Use the predictions dataset as the pool of a [labeling flow](/use-cases/organize-labeling), so labelers correct boxes instead of drawing them.
* **Score it.** If you have a few labeled images, measure it properly with [Evaluate and compare models](/use-cases/evaluate-and-compare).
* **Fine-tune it.** When its vocabulary isn't yours, [train a detector](/use-cases/train-a-detector) on your own labels.