| Labeling your images | AI-drafted, Auto Label labels the classes you describe in a text prompt; Smart Polygon powered by Meta AI's Segment Anything; Label Assist pre-annotates with your own trained models | By hand, image-level labels and bounding boxes drawn in the console, or import a SageMaker Ground Truth manifest |
| Improving a model after launch | Built in, active learning and operator feedback send flagged production images back into your dataset, and Vision Events logs every prediction and image | Manual, add and label new images, then train a new model version, or install the separate Model Feedback sample solution from GitHub |
| Running a trained model | Ready when you call it, serverless hosted API that scales with traffic, or self-hosted on your own hardware, with no model to start or stop | Start and stop each version, a model must be started before it can analyze images; inference units autoscale to a set maximum and bill per hour while running, even when idle |
| Ready-made models | 1,000,000+ datasets and 250,000+ pre-trained models on Roboflow Universe | Yes, pre-trained APIs for labels, faces, face liveness, text, unsafe content, celebrities, and PPE |
| Vision tasks you can train | Object detection, instance and semantic segmentation, classification, keypoints, OCR, depth estimation, and multimodal tasks | Classification and detection only; AWS says Custom Labels is not designed for faces, text, or unsafe content |
| Choose your model architecture | Yes, RF-DETR, YOLO (v8 / v11 / v26), Detectron2, and more | Not documented; Custom Labels trains an AWS-managed model |
| Export and own model weights | Yes | Not documented; models are called through the Custom Labels API |
| Data formats and limits | 40+ formats, including COCO, YOLO, and Pascal VOC; dataset versioning, preprocessing, and augmentation built in | PNG and JPEG only; up to 250 labels per dataset, 50 bounding boxes per image, and 4096 x 4096 pixels per image |
| Edge and offline deployment | Yes, NVIDIA Jetson, GPU servers, Raspberry Pi, x86, or Roboflow AI1, online or fully offline | AWS cloud only; Custom Labels models run on inference units in an AWS Region; edge or offline use not documented |
| Video and camera streams | Any RTSP, USB, GigE, or industrial camera | Stored video, analyzed from Amazon S3; Streaming Video closed to new customers as of April 30, 2026 |
| Sub-agents for frontier models (MCP) | Yes, every trained model is exposed over MCP so frontier models can call it as a specialized vision tool | Not documented |
| Uptime SLA | Yes, on Enterprise, with a public status page | Yes, 99.9% monthly uptime per AWS Region, with service credits |
| Compliance | SOC 2 Type II, HIPAA with BAAs, PCI DSS | Assessed under AWS programs including SOC, PCI, FedRAMP, and HIPAA |
| Self-hosted inference | Yes, open-source inference server for custom and foundation models | No; Custom Labels models are served as AWS-managed endpoints billed in inference units |
| Open-source CV library | Yes, supervision, with model-agnostic utilities for detections, tracking, annotation, and evaluation | None comparable in mainstream community use |
| SDKs and APIs | Yes, Python SDK plus REST/HTTP APIs | Yes, AWS SDKs (boto3 for Python, plus Java, .NET, Go, JavaScript, Ruby, PHP) and the AWS CLI, tied to AWS IAM authentication |
| Free tier | Yes, with free self-hosted inference, model weights download, and a private workspace | Pay as you go; Custom Labels $1 per training hour and $4 per inference hour |
| Pricing | Published plans; Enterprise adds volume pricing and an uptime SLA | Image APIs from $0.001 per image for the first 1 million images a month; Custom Labels $1 per training hour and $4 per inference hour |