Resurrecting Yahoo’s discontinued NSFW model in Keras 3

What it does

This is a Keras reimplementation of Yahoo’s discontinued Open-NSFW model, which was originally built in the now-obsolete Caffe framework. It focuses specifically on detecting pornographic images and videos, wrapping the old weights in a simple high-level API for single images, batches, and video frames. Lower-level primitives are also exposed if you want to fine-tune the model or build your own pipeline.

The interesting bit

The project preserves Yahoo’s original preprocessing pipeline—including an optional JPEG round-trip that matches the Caffe implementation—alongside a simpler modern alternative, and it auto-downloads the original pre-trained weights so the model works out of the box. It targets Keras 3 with both TensorFlow and JAX backends, though the README candidly notes that PyTorch support is an afterthought with divergent output.

Key highlights

• Supports both Keras 3 (TensorFlow or JAX) and legacy tf-keras.
• End-to-end APIs for images, image batches, and video frame analysis with configurable aggregation.
• Can generate Grad-CAM overlays to highlight prediction hotspots (TensorFlow backend only).
• Automatically fetches pre-trained Yahoo weights on first use.
• Video processing includes frame skipping, probability aggregation, and optional annotated output.

Caveats

• PyTorch backend works but produces noticeably different inference results and runs slower; the README explicitly advises against using it.
• Grad-CAM visualization is only available when using the TensorFlow backend.
• The model is narrowly scoped to pornographic content and does not claim to detect other forms of NSFW material.

Verdict

Developers who need a drop-in, offline NSFW classifier for images or video should look here; those seeking a general-purpose content moderation tool or a PyTorch-native workflow should probably keep looking.