Skip to main content
Back

Falcons.ai's Open-Source NSFW Detector Hits 4M Downloads Without a Single API

A fine-tuned Vision Transformer for binary NSFW classification has crossed 4 million downloads on Hugging Face, becoming the default open-source content moderation baseline.

  • Falconsai/nsfw_image_detection has passed 4 million downloads on Hugging Face.
  • Binary classifier fine-tuned from google/vit-base-patch16-224-in21k on 80,000 curated images.
  • Reports 98.04% accuracy on the maintainer's internal evaluation set, no public benchmark.
  • Apache 2.0 license, runs via a three-line Transformers pipeline call, Safetensors and PyTorch weights.
  • Used in production moderation pipelines and academic studies of AI-generated harmful content.
  • Competing 384x384 variant from AdamCodd trades speed for higher-resolution detail.

An open-source image classifier from Falcons.ai has quietly become one of the most downloaded content moderation models on Hugging Face, with more than 4 million downloads and over 1,100 likes. Falconsai/nsfw_image_detection is a fine-tuned Vision Transformer that does one thing: sort images into normal or nsfw. It ships under Apache 2.0, runs on a single GPU, and has become a common baseline for teams that need a safety filter without paying per API call.

A small ViT with one job

The checkpoint is a fine-tune of Google's ViT-Base backbone, specifically google/vit-base-patch16-224-in21k, which was pre-trained on ImageNet-21k. Falconsai further trained it on a proprietary set of 80,000 images labeled normal and nsfw. Inputs are standard 224x224 RGB images, and outputs are two logits you can softmax into a score.

Training was deliberately vanilla: batch size 16, learning rate 5e-5, no exotic tricks. The maintainer reports 98.04% accuracy on an internal test set. No independent benchmark or evaluation on public datasets exists, which makes it hard to compare against competing NSFW detectors or gauge robustness under adversarial inputs.