Image Analysis Compute

Neural network work (both training a custom model on your annotations and running any model on new images) runs on cloud servers, not in your browser. Those server runs are what consume compute usage in image analysis.

What Consumes Compute Usage

Two server-side operations consume usage:

• Training a custom model. When you finish labeling examples in Click-Detect, you press Train to build a U-Net detector from your annotations. Training runs on a GPU server.
• Running inference on an image. Each time you run a trained model on an image, detection runs on the server.

Inference is per-image. Running a model across multiple images means multiple jobs and the usage adds up accordingly. The estimated usage is shown in the run dialog box before you confirm.

What's Always Free

Every other image-analysis operation runs in your browser for free:

• Viewing and navigating images with pan, zoom, Z-stack, timelapse, and presentation
• Channel controls and enhancements for fluorescence images, including composites and look-up tables
• Manual annotations like polygons, points, lines, scale bars, and measurements
• Cropping, regions of interest, and overlays
• Editing detections by adding, removing, or adjusting individual results after a run
• Confidence threshold adjustments on a saved result set (filtering happens in the browser)
• All exports, including annotated images, detection tables, and model cards

The Workflow

The workspace stepper at the top of the screen walks you through three phases: Labeling, Training, and Testing.

Labeling

1. Open an image in the workspace and enter the Labeling phase
2. The image is divided into a four-quadrant grid. Select a quadrant to focus on that region
3. Click on objects to label them. Use the Click, Box, or Move tools in the sidebar
4. Accept each detection, then Mark Complete when you've labeled every object in the quadrant
5. Repeat for all four quadrants, then move to the next image. The sidebar tracks your progress (sections completed, total objects labeled)

Training

1. Once you have enough labels, click Review & Train in the sidebar
2. The review screen shows all your labeled objects across images. Remove any bad labels before proceeding
3. Click Start Training. Training runs on a GPU server and progress shows in real time (preparing, uploading, queued, training, completed)
4. When training finishes, the model is saved as a Model Card for your project

Testing

1. After training completes, move to the Testing phase
2. Pick an image and click Run to run your trained model on it
3. Adjust the confidence threshold to filter results. That step is instant and free

Tracking Jobs

Server jobs appear in your team's Compute Jobs list (Settings → Usage → Compute Jobs). Each row shows job type, who started it, status, duration, and usage. You can also export the list as CSV.

Next Steps

• Compute usage for how compute usage is pooled at the team level
• Image AI and data for the full guide to training and running detection
• Analyze images for the overall image-analysis workflow