How do I add a new class to an existing computer vision model without losing my labels?

Key facts

What you need

Your labeled dataset, plus images that show the new class

What you get

A new model version that finds the old classes and the new one

Image sources

Your existing project, new uploads and Roboflow Universe

Runs on

Cloud API, or your own device or server running Docker

How to add a new class, step by step

Six steps from your current dataset to a model that knows one more class.

1. 01

   Connect Roboflow to your coding agent, ChatGPT or Codex

   Add the Roboflow MCP server to Claude, Cursor or Gemini CLI, or install the Roboflow plugin in ChatGPT or Codex. Your assistant can then upload images, auto-label the new class, start training and test the new model from a prompt, at no extra cost beyond your Roboflow plan.

2. 02

   Upload images that show the new class to the same project

   Add images of the new class to the project that holds your current dataset, so every existing label stays in place. Use varied examples of the new class, taken in the same conditions the model will see.

3. 03

   Add the new class to the images you already labeled

   Run Auto Label in Roboflow Annotate with SAM 3 and a text prompt that describes the new class. When it asks what to do with existing annotations, choose to keep or merge them, not replace them, then review the labels it added.

4. 04

   Label every object on the new images

   On the images you just uploaded, label the existing classes as well as the new one. New images with only the new class labeled would teach the model to ignore the classes it already knows.

5. 05

   Create a new dataset version and train

   Generate a new dataset version. A version is a snapshot of your images and labels, so the earlier version stays as it was. Train an RF-DETR model on the new version with Roboflow Train.

6. 06

   Check the new model finds every class

   Run the new model on a test image that contains both the old classes and the new one, and count what it finds by class name.

   PythonCopy

   import supervision as sv
   from collections import Counter
   from inference import get_model
   
   model = get_model("your-project/2")
   
   result = model.infer("test_image.jpg", confidence=0.4)[0]
   detections = sv.Detections.from_inference(result)
   
   print(Counter(detections.data["class_name"]))
   # Counter({'hard-hat': 5, 'vest': 4, 'gloves': 3})

   Rather not write code? Ask the Roboflow agent to set this up for you →

Why 2 million engineers build on Roboflow

• Label faster

  Auto-label with SAM 3 or your own model, then fix the misses instead of drawing every box.

• RF-DETR

  Roboflow’s open-source real-time detection model, built for accuracy (opens in a new tab) on small objects.

• Logic built in

  Counting, zones, tracking and alerts are ready-made Workflow blocks, so there is less code to write.

• Your model, your weights

  Use your trained weights on every plan, including free, through the open-source Inference package.

• Run it anywhere

  Self-host on a workstation, Docker, an on-prem GPU or an edge device, or call the cloud API.