Superb Curate

Intelligent Data Curation Made Easy

AI-based product and tooling to manage all your computer vision datasets in one place, curate your most valuable data, optimize labeling spend, and build better models.
curate-Data Management

Data Management

upload and pipeline large volumes of data as it’s collected


Raw Data





curate-Visualization & Analytics

Visualization & Analytics

to evaluate similarity and distribution


Use embeddings to understand distribution by visual similarity.


Use analytics to see distribution by metadata and annotations.

curate-Query & Slice

Query & Slice

using metadata and annotations


Easily find data matching your desired conditions with query.


Group meaningful data into distinct slices.

curate-Auto Curation

Auto Curation

with high-quality embeddings


Auto-curate a well-balanced training or validation dataset.


Detects edge cases and mislabels in your dataset.

“Carefully balancing resources for unproven ML projects with the amount of quality data needed for commercialization was tough, but Superb Curate changed that. It auto-curates impactful data to label first and combines that with powerful search, allowing us to add new object classes significantly faster. It also reduces retraining time, as it proved capable of uniformly extracting teaching data from the feature map, enabling data reduction without feature bias.”

Nippon Steel Corporation l Nobuyuki Tatemizo


Achieve better model performance 
with less data


Better Models

Improve precision and recall by 15% or more across object classes without collecting additional data (compared to random sampling).


Less Labeling Costs

Create more robust models using only 25% of the typical data needed without reducing performance (0.05% or less change in F-1 across object classes).


Unified Platform

Scale better together by empowering your engineers, labelers, and data team to collaborate on one platform, increasing development speed and reducing friction.

Key Features

Auto-generates high-quality embeddings whenever new images or objects are uploaded, so you can cluster data without manual curation or custom models.

Seamlessly upload and pipeline your data, including images with associated annotations and metadata, into a single platform as soon as it's collected.

Easily find and group data that meets your desired conditions using any combination of metadata or annotation information tagged within your images.

Automatically curate the most suitable dataset for your model needs, including curating a balanced training or validation set, detecting edge cases, or uncovering mislabels.

Visualize the distribution of images or objects, clustered based on visual similarities, over a two-dimensional space to understand patterns in your dataset and find outliers.

Get deep insight into trends and patterns within your data by analyzing the distribution of metadata, annotation types, object classes, and more.

Integrates seamlessly with Superb Label. Identify the highest value data to label first, send a labeling batch to Label, and resend to Curate in one click.

Perform powerful embedding-based searches using natural language or image query to find the exact data you need within your own or open-source datasets.

Upload model predictions to visually compare inference results to ground truths, identify underperforming slices, and curate data that are most likely to improve your model.

“The addition of AI-driven DataOps features to the Superb Suite moves the platform even more securely into an integrated data management solution for computer vision data. The ability to curate datasets to find suggested changes will be very powerful for us.”

Blaine Bateman, Chief Data Scientist

How Auto-Curate Works

Superb Curate provides proprietary, high-dimensional embedding generation algorithms and uses unsupervised learning to cluster image and object data on visual similarity. Based on these clusters, our curation algorithm automatically selects the data most suitable for your model needs.

1. Upload Data

Improve precision and recall by 15% or more across object classes without collecting additional data (compared to random sampling).


2. Select and Run

Select curation parameters by choosing a name, the number of images you want to curate, and, for object curation, the classes you wish to target.


3. Review Results

Curate automatically saves your output as a new slice and provides deep analytics and reporting so you can review criteria and data distribution in detail.

Curation Methods

Curates a balanced slice that best represents your entire dataset. Suitable for an initial training set.

Curates a high-quality validation set that you can use to validate model performance.

Prioritizes images or objects that are sparsely located and likely edge cases.

Uncovers and prioritizes data that is most likely to be mislabeled.

curate_Curate what to label

These teams build datasets engineers want with Superb AI


Give it a try

Join 5,000 machine learning teams already using Superb AI.