Principle:NVIDIA NeMo Curator NSFW Content Filtering
| Metadata | |
|---|---|
| Knowledge Sources | N/A |
| Domains | Data_Curation, Image_Processing, Content_Safety |
| Last Updated | 2026-02-14 |
Overview
NSFW Content Filtering is a technique for detecting and removing not-safe-for-work content from image datasets using a classifier applied to CLIP embeddings.
Description
NSFW Content Filtering in NeMo Curator employs a binary classifier operating on CLIP image embeddings to predict the probability that an image contains not-safe-for-work content. The classifier assigns each image an NSFW probability score, and images with scores exceeding a configurable threshold are removed from the dataset. This approach enables efficient content safety filtering at scale because the classifier operates on lightweight embedding vectors rather than raw pixel data, and the CLIP embeddings capture the semantic content necessary for accurate NSFW detection.
Usage
Use NSFW Content Filtering after the CLIP Embedding stage to remove inappropriate or unsafe content from image datasets. This stage is critical for any dataset that will be used in production systems, public-facing applications, or model training where content safety is a requirement. Apply this filter before exporting curated datasets to ensure compliance with content policies.
Theoretical Basis
NSFW Content Filtering is based on training a logistic regression or shallow MLP classifier on CLIP image embeddings using labeled NSFW and SFW (safe-for-work) training data. CLIP embeddings encode rich semantic information about image content, including the ability to distinguish between safe and unsafe visual content. The classifier learns a decision boundary in the CLIP embedding space that separates NSFW content from safe content. Because CLIP was trained on a diverse corpus of image-text pairs from the internet, its embeddings capture nuanced visual concepts relevant to content safety classification. The binary nature of the classifier produces a probability score between 0 and 1, where higher scores indicate greater likelihood of NSFW content. By setting a threshold on this probability, users can control the sensitivity of the filter, trading off between false positives (safe images incorrectly flagged) and false negatives (NSFW images that pass through).