Inference#

Ultralytics Platform provides browser-based inference for testing trained models and dedicated endpoints for programmatic access.

Predict Tab#

Every model with weights includes a Predict tab for browser-based inference:

1. Navigate to your model
2. Click the Predict tab
3. Upload an image, use an example, or open your webcam
4. Review the task-specific overlay, prediction summary, timing, and raw response

Models without weights show an empty state instead — train the model or upload weights first.

Input Methods#

The predict panel supports multiple input methods:

graph LR
    A[Upload Image]:::start --> D[Auto-Inference]:::proc
    B[Example Image]:::start --> D
    C[Webcam Capture]:::start --> D
    D --> E[Results + Overlays]:::out

    classDef start fill:#4CAF50,color:#fff
    classDef proc fill:#2196F3,color:#fff
    classDef out fill:#9C27B0,color:#fff

Upload Image#

Drag and drop or click to upload:

• Supported formats: JPEG, PNG, WebP, AVIF, HEIC, JP2, TIFF, BMP
• Max size: 10 MB
• Auto-inference: Results appear automatically after upload

Example Images#

The predict panel shows up to two example images from your model's linked dataset, preferring the val split, then test, then train. If no dataset is linked, default examples are used:

For OBB models, aerial images of boats and an airport are shown instead.

Webcam#

Select Webcam above the image area to start a live camera feed:

1. Grant camera permission when prompted
2. Click the video preview to capture a frame
3. Inference runs automatically on the captured frame
4. Click Back to webcam to return to the live feed

On your own deployment's Predict tab, the webcam runs inference continuously instead. See Live Camera Inference.

View Results#

Inference results display the output appropriate to the model task: boxes, masks, keypoints, oriented boxes, classification scores, semantic coverage, or a depth map. Object results use the dataset class colors when available. The panel also shows preprocess, inference, postprocess, and network timing.

The results panel shows:

Two controls sit over the preview once results are in: click the image to enlarge it with overlays intact, and use the download button to save an annotated JPEG of the current result.

Inference Parameters#

Adjust inference behavior with the three sliders below the image (depth models show only Image Size):

Confidence Threshold#

Filter predictions by confidence:

• Higher (0.5+): Fewer, more certain predictions
• Lower (0.1-0.25): More predictions, some noise
• Default (0.25): Balanced for most use cases

IoU Threshold#

Control Non-Maximum Suppression:

• Higher (0.7+): Allow more overlapping boxes
• Lower (0.3-0.5): Suppress overlapping detections more aggressively
• Default (0.7): Balanced NMS behavior for most use cases

Deployment Predict#

Each running dedicated endpoint includes a Predict tab on its deployment page. This uses the deployment's own inference service rather than the shared predict service, letting you test your deployed endpoint from the browser.

On a ready endpoint, processed images also contribute to the Monitoring tab. Its examples and aggregate charts are lightweight, temporary data held in memory; stopping, restarting, redeploying, resizing, or replacing the model can clear them. Save examples to a dataset to keep them.

Live Camera Inference#

On the Predict tab of a deployment you own, select Webcam or IP camera to run the endpoint on live video:

Live inference uses the endpoint's bound API key, which only the workspace owner can load; for other team members the webcam captures single frames and IP camera is unavailable, as on a model's Predict tab. The IP camera must be reachable from the internet: the endpoint refuses local network addresses such as 192.168.x.x. Each result is for the newest frame, so frames are skipped when inference falls behind. Slider changes apply to the next webcam frame and restart an IP camera stream. Live inference pauses while the browser tab is hidden. Click the preview to capture a frame, or Disconnect to stop viewing the IP camera.

Background Camera#

An endpoint with a custom CPU and memory size can keep watching one IP camera after you disconnect or close the page. Once the connected camera shows results, turn on Keep running in the background. The deployment header shows Camera on, and results go to the Monitoring tab as temporary examples and prediction statistics.

• Settings: The background camera always uses the default confidence (0.25), IoU (0.7), and the model's training image size; the sliders do not apply to it.
• Cost: It runs on the endpoint's warm instance at no extra charge; the hourly uptime rate applies whether the camera is on or off.
• Changes: Turning the camera on, off, or to another camera restarts the endpoint's instance, which keeps the endpoint ready but clears its temporary monitoring data.
• Stopping: Turn the switch off. Disconnecting or closing the page does not stop it, and resizing the endpoint to the default size removes it. If the camera goes offline, the endpoint keeps reconnecting.
• Endpoint lifecycle: Stopping the endpoint stops the camera and the charges; starting it again resumes the saved camera.

Default-size endpoints offer live webcam and IP camera inference without the background option. To save a background camera from the API, use the deployment camera action.

Stream Results from the API#

Send an RTSP or RTSPS URL as source with the Accept: text/event-stream header to a dedicated endpoint URL to receive results as server-sent events:

curl -N -X POST \
  "https://YOUR_DEPLOYMENT_URL.run.app/predict" \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Accept: text/event-stream" \
  -F "source=rtsp://user:password@camera.example.com:554/stream" \
  -F "conf=0.25"

Each frame event carries images in the response shape with normalized (0-1) coordinates, a preview JPEG data URL of the frame, and metadata with the task and class names. Only conf, iou, and imgsz apply, and streaming the endpoint's background camera URL uses its default settings. Events with only a status carry no frame, and an event with an error message (the camera could not be read, or the endpoint cannot run the model) ends the stream. The stream also closes when the endpoint restarts or the request reaches its time limit, so reconnect with a backoff when it ends without an error. The Platform API's deployment predict route and SDK do not stream; send camera requests to the endpoint URL with its bound API key. A camera source without the header returns 400.

Dedicated Endpoint API#

The API Docs card in the model Predict tab contains example Python, JavaScript, and cURL requests, pre-filled with the confidence, IoU, and image size currently set on the sliders. The URL and key are placeholders until you deploy the model — a Deploy button next to the code tabs jumps to the model's Deploy tab. After deployment, the Docs result tab in the deployment page's Predict tab fills in that endpoint's URL and, for workspace owners, its bound API key, ready to copy and run.

Authentication#

Include your API key in requests:

Authorization: Bearer YOUR_API_KEY

Endpoint#

Dedicated endpoints take requests on their own URL:

POST https://YOUR_DEPLOYMENT_URL.run.app/predict

Shared inference uses the Platform API with the model's full path:

POST https://platform.ultralytics.com/api/models/{owner}/{project}/{model}/predict

Both accept the same multipart/form-data body and return the same response shape. With the Python SDK, use client.models.predict(owner, project, model, body=...) for shared inference or client.deployments.predict(owner, deployment, body=...) for a dedicated deployment. Both SDK methods call the Platform API, so its rate limits and request size limit apply. To avoid them, post directly to a dedicated endpoint URL as shown under Request. Shared inference example:

from ultralytics_platform import Platform

client = Platform()  # reads ULTRALYTICS_API_KEY
with open("image.jpg", "rb") as f:
    results = client.models.predict("acme-vision", "inspection", "v3", body={"file": f, "conf": 0.25})

Request#

import requests

url = "https://YOUR_DEPLOYMENT_URL.run.app/predict"
headers = {"Authorization": "Bearer YOUR_API_KEY"}
data = {"conf": 0.25, "iou": 0.7, "imgsz": 640}

with open("image.jpg", "rb") as image_file:
    response = requests.post(url, headers=headers, files={"file": image_file}, data=data)
print(response.json())

Request Parameters#

Response#

{
    "images": [
        {
            "shape": [1080, 1920],
            "results": [
                {
                    "class": 0,
                    "name": "person",
                    "confidence": 0.92,
                    "box": { "x1": 100, "y1": 50, "x2": 300, "y2": 400 }
                },
                {
                    "class": 2,
                    "name": "car",
                    "confidence": 0.87,
                    "box": { "x1": 400, "y1": 200, "x2": 600, "y2": 350 }
                }
            ],
            "speed": {
                "preprocess": 1.2,
                "inference": 12.5,
                "postprocess": 2.3
            }
        }
    ],
    "metadata": {
        "imageCount": 1,
        "classNames": ["person", "bicycle", "car", "..."],
        "functionTimeAlive": 1284.51,
        "functionTimeCall": 0.018,
        "task": "detect",
        "version": {
            "ultralytics": "8.x.x",
            "torch": "2.6.0",
            "torchvision": "0.21.0",
            "python": "3.13.0"
        }
    }
}

Response Fields#

Task-Specific Responses#

Response format varies by task:

{
  "class": 0,
  "name": "person",
  "confidence": 0.92,
  "box": {"x1": 100, "y1": 50, "x2": 300, "y2": 400}
}

Rate Limits#

The shared model API is limited to 20 requests/minute for each API key, signed-in caller, or anonymous IP. The Platform deployment predict route (POST /api/deployments/{owner}/{deployment}/predict) has the same limit. When throttled, the API returns 429 with a Retry-After header. See the full rate-limit reference for all endpoint categories.

Error Handling#

Common error responses:

FAQ#