# What's in the Viam registry for vision

Three kinds of registry entries feed a vision pipeline: ML model service implementations, vision service models, and public ML models. How to pick among them.
> Source: https://docs.viam.com/vision/deploy-and-maintain/available-models/


The [Viam registry](https://app.viam.com/registry) has three kinds of entries you reach for when building a vision pipeline. Understanding what each one is makes the rest of the deploy and tune flow easier.

## Three kinds of entries

- **ML model service implementations.** The ML model service has one API method, `Infer`, but the registry offers multiple module implementations of that service. Each implementation targets a specific framework: TFLite, ONNX, TensorFlow, or PyTorch. Some also target specific hardware like CPU or GPU on Jetson. When you add an ML model service to your machine, you pick one implementation.
- **Vision service models.** The vision service has one API that covers detection, classification, and 3D segmentation (see the [full API reference](/reference/apis/services/vision/)). Multiple models implement that API. Three ship built into Viam: `mlmodel`, `color_detector`, and `viam:vision:detections-to-segments`. Registry modules add more for specialized tasks. You pick a vision service model when you configure a vision service.
- **Public ML models.** Model artifacts (the weights plus metadata) published to the registry so any machine can deploy them without training. Your ML model service loads one of these.

A typical pipeline uses one of each: a camera → an ML model service implementation running a public ML model → a vision service model interpreting its output.

## How to pick

The registry mixes Viam-authored modules, partner contributions, and community work at various stages of maturity. Before picking one:

- **Read the module's README** for tensor requirements, hardware support, and example configuration. The registry card links to it.
- **Check recent commits and versions** if maturity matters. A module with a single 2023 release is a different bet than one with continuous releases this quarter.
- **Prefer framework-specific implementations over the generic `triton`** unless you actually need GPU or multi-framework flexibility. Framework-specific implementations are simpler to configure and debug.
- **Verify the model task type matches your need.** A model tagged for `object detection` will not give you classifications, and vice versa. See [Deploy an ML model from the registry, step 1](/vision/deploy-and-maintain/deploy-from-registry/#1-pick-a-model) for the task/framework/hardware matching table.

To browse or search the registry, go to [app.viam.com/registry](https://app.viam.com/registry).

## Next steps

- [Deploy an ML model from the registry](/vision/deploy-and-maintain/deploy-from-registry/): the step-by-step how-to
- [Deploy a custom ML model](/vision/deploy-and-maintain/deploy-custom-model/): for models you trained yourself or brought from elsewhere
- [Configure a vision pipeline](/vision/configure/): wire a model through an ML model service and a vision service