From the model card
Fields below are copied from the tags and counters on the HuggingFace repository Falconsai/nsfw_image_detection at our last fetch. They are set by the uploader, not verified by us; rows with no tag are omitted. How this page is made.
- Publisher (HF namespace)
- Falconsai
- Pipeline tag
- image-classification
- Library
- Transformers
- Framework tags
- PyTorch
- Weight formats
- safetensors
- License tag
apache-2.0— read the license file in the repo before relying on it- Papers cited
- arXiv:2010.11929
- Downloads (HF counter at last fetch)
- 4,031,803
- Likes (HF counter at last fetch)
- 1,170
- Model card
- https://huggingface.co/Falconsai/nsfw_image_detection
Use cases
- Automated content moderation in user-generated image platforms
- Pre-screening uploads before expensive human review
- Filtering image datasets for safety before model training
- Enforcing content policies at ingestion points of image-accepting APIs
- First-pass flagging layer upstream of more granular classifiers
Pros
- Single-purpose binary classification simplifies deployment logic
- ViT architecture handles compositional and varied image content
- Apache 2.0 license; available for CPU inference
- Zero labeled data required for deployment vs. training from scratch
Cons
- Binary safe/unsafe classification misses nuanced harmful content categories (violence, gore, self-harm)
- Edge cases — medical imagery, classical art, partial exposure — regularly misclassified
- Training dataset provenance not publicly disclosed, limiting auditing
- Probability scores are not calibrated explanations — no rationale output
- Requires calibration and threshold tuning before production content moderation
Tags
transformerspytorchsafetensorsvitimage-classificationarxiv:2010.11929license:apache-2.0endpoints_compatibleregion:usdeploy:azurenot-for-all-audiences