About Image Curation#
Learn how to curate high-quality image datasets using NeMo Curator’s powerful image processing pipeline. NeMo Curator enables you to efficiently process large-scale image-text datasets, applying quality filtering, content filtering, and semantic deduplication at scale.
Use Cases#
Prepare high-quality image datasets for training generative AI models such as LLMs, VLMs, and WFMs
Curate datasets for text-to-image model training and fine-tuning
Process large-scale image collections for multimodal foundation model pretraining
Apply quality control and content filtering to remove inappropriate or low-quality images
Generate embeddings and semantic features for image search and retrieval applications
Remove duplicate images from large datasets using semantic deduplication
Architecture#
NeMo Curator’s image curation follows a modular pipeline architecture where data flows through configurable stages. Each stage performs a specific operation and passes processed data to the next stage in the pipeline.
flowchart LR
A[Tar Archive Input] --> B[File Partitioning]
B --> C[Image Reader<br/>DALI GPU-accelerated]
C --> D[CLIP Embeddings<br/>ViT-L/14]
D --> E[Aesthetic Filtering<br/>Quality scoring]
E --> F[NSFW Filtering<br/>Content filtering]
F --> G[Duplicate Removal<br/>Semantic deduplication]
G --> H[Export & Sharding<br/>Tar + Parquet output]
classDef input fill:#e1f5fe,stroke:#0277bd,color:#000
classDef processing fill:#f3e5f5,stroke:#7b1fa2,color:#000
classDef output fill:#e8f5e8,stroke:#2e7d32,color:#000
class A input
class B,C,D,E,F,G processing
class H output
This pipeline architecture provides:
Modularity: Add, remove, or reorder stages based on your workflow needs
Scalability: Distributed processing across multiple GPUs and nodes using Ray
Flexibility: Configure parameters for each stage independently
Efficiency: GPU-accelerated processing with DALI and CLIP models
Introduction#
Master the fundamentals of NeMo Curator’s image curation pipeline and set up your processing environment.
Learn about ImageBatch, ImageObject, and pipeline stages for efficient image curation
Learn prerequisites, setup instructions, and initial configuration for image curation
Curation Tasks#
Load Data#
Load and process large-scale image datasets from local storage using tar archives with GPU-accelerated DALI for efficient distributed processing.
Load and process JPEG images from tar archives using DALI
Process Data#
Transform and enhance your image data through classification, embeddings, and filters.
Apply built-in filters for aesthetic quality and NSFW content filtering.
Generate image embeddings using CLIP models.
Pipeline Management#
Optimize and manage your image curation pipelines with advanced execution backends and resource management.
Configure Ray-based executors for distributed processing and resource management.
Optimize performance with DALI GPU acceleration and efficient resource allocation.
Save & Export#
Export your curated image datasets with metadata preservation, custom resharding options, and support for downstream training pipelines.
Save metadata to Parquet and export filtered datasets with custom resharding.