Overview | NeMo Curator

Learn what video curation is and how you use NeMo Curator to turn long videos into high‑quality, searchable clips. Depending on the use case, this can involve processing 100+ PB of videos. To efficiently process this quantity of videos, NeMo Curator provides highly optimized curation pipelines.

Use Cases

Identify when to use NeMo Curator by matching your goals to common video curation scenarios.

Generating clips for video world model training
Generating clips for generative video model fine-tuning
Creating a rich video database for video retrieval applications

Architecture

Understand how components work together so you can plan, scale, and troubleshoot video pipelines. The following diagram outlines NeMo Curator’s video curation architecture:

High-level outline of NeMo Curator's video curation architecture

Video pipelines use the XennaExecutor backend by default, which provides optimized support for GPU-accelerated video processing including hardware decoders and encoders. You do not need to import or configure the executor unless you want to use an alternative backend. For more information about customizing backends, refer to Pipeline Execution Backends.

Introduction

Get oriented and prepare your environment so you can start curating videos with confidence.

Concepts

Learn about the architecture, stages, pipelines, and data flow for video curation stages pipelines ray

Get Started

Install NeMo Curator, configure storage, prepare data, and run your first video pipeline.

Curation Tasks

Follow task-based guides to load, process, and write curated video data end to end.

Load Data

Bring videos into your pipeline from local paths or remote sources you control.

Local & Cloud

Load videos from local paths or S3-compatible and HTTP(S) URLs. local s3 file-list

Remote (JSON)

Provide an explicit JSON file list for remote datasets under a root prefix. file-list s3

Process Data

Transform raw videos into curated clips, frames, embeddings, and metadata you can use.

Clip Videos

Split long videos into shorter clips using fixed stride or scene-change detection. clips fixed-stride transnetv2

Encode Clips

Encode clips to H.264 using CPU or GPU encoders and tune performance. clips h264_nvenc libopenh264 libx264

Filter Clips and Frames

Apply motion-based filtering and aesthetic filtering to improve dataset quality. clips frames motion aesthetic

Extract Frames

Extract frames from clips or full videos for embeddings, filtering, and analysis. frames nvdec ffmpeg fps

Create Embeddings

Generate clip-level embeddings with Cosmos-Embed1 for search and duplicate removal. clips cosmos-embed1

Create Captions & Preview

Generate Qwen‑VL captions and optional WebP previews; optionally enhance with Qwen‑LM. captions previews qwen webp

Remove Duplicate Embeddings

Remove near-duplicates using semantic clustering and similarity with generated embeddings. clips semantic pairwise kmeans

Write Data

Save outputs in formats your training or retrieval systems can consume at scale.

Save & Export

Understand output directories, parquet embeddings, and packaging for training. parquet webdataset metadata

Tutorials

Practice with guided, hands-on examples to build, customize, and run video pipelines.

Beginner Tutorial

Create and run your first video pipeline: read, split, encode, embed, write. splitting encoding embeddings

Pipeline Customization Tutorials

Customize environments, code, models, and stages for video pipelines. environments code models stages