nsfw image detector

---
license: mit
base_model:

• timm/eva02_base_patch14_448.mim_in22k_ft_in22k_in1k

• pytorch

• transformers

EVA-based Fast NSFW Image Classifier

Table of Contents

• Model Description

• Try it Online!

• Model Performance Comparison

• Usage

• Training

• Speed and Memory Metrics

Model Description

This model is a vision transformer based on the EVA architecture, fine-tuned for NSFW content classification. It has been trained
to detect four categories (neutral, low, medium, high) of visual content using 100,000 synthetically labeled images.

The model can be used as a binary (true/false) classifier if desired, or you can obtain the full output probabilities.. It outperforms other excellent publicly available models such as Falconsai/nsfw_image_detection or AdamCodd/vit-base-nsfw-detector in our internal benchmarks adding the enrichment of being able to select the NSFW level that suits your use case.

Try it Online! 🚀

You can try this model directly in your browser through our Hugging Face Space. Upload any image and get instant NSFW classification results without any installation required.

Model Performance Comparison

Global Performance

| Category | Freepik | Falconsai | Adamcodd |
|----------|-------------|------------------|----------------|
| High | 99.54% | 97.92% | 98.62% |
| Medium | 97.02% | 78.54% | 91.65% |
| Low | 98.31% | 31.25% | 89.66% |
| Neutral | 99.87% | 99.27% | 98.37% |

In the table below, the results are obtained as follows:

• For the Falconsai and AdamCodd models:

• For the Freepik model:

Conclusions:

• Our model outperforms AdamCodd and Falconsai in accuracy. It is entirely fair to compare them on the "high" and "neutral" labels.

• Our model offers greater granularity. It is not only suitable for detecting "high" and "neutral" content, but also performs excellently at identifying "low" and "medium" NSFW content.

Accuracy by AI Content

We have created a manually labeled dataset with careful attention to avoiding biases (gender, ethnicity, etc.). While the sample size is relatively small, it provides meaningful insights into model performance across different scenarios, which was very useful in the training process to avoid biases.

The following tables show detection accuracy percentages across different NSFW categories and content types:

#### AI-Generated Content

| Category | Freepik Model | Falconsai Model | Adamcodd Model |
|----------|-------------|------------------|----------------|
| High | 100.00% | 84.00% | 92.00% |
| Medium | 96.15% | 69.23% | 96.00% |
| Low | 100.00% | 35.71% | 92.86% |
| Neutral | 100.00% | 100.00% | 66.67% |

Conclusions:

• Avoid using Falconsai for AI-generated content to prevent prediction errors.

• Our model is the best option to detect NSFW content in AI-generated content.

Usage

Quick Start via pip

```sh
pip install nsfw-image-detector

Initialize the detector

Load and classify an image

Check if the image contains NSFW content sentivity level medium or higher

Get probability scores for all categories

Example output:

### Quick Start with Pipeline

Create classifier pipeline

Load and classify an image

Example output:

The model supports efficient batch processing for multiple images:

Note: If the intention is to use the model in production review Speed and Memory Metrics section before using this approach.
Avoid installation of pip dependency
The following example demonstrates how to customize the NSFW detection label, it is very similar to the code in PyPy. This code returns True if the NSFW level is 'medium' or higher:

device = "cuda" if torch.cuda.is_available() else "cpu"

Load model and processor

Load original processor (faster for tensors)

def predict_batch_values(model, processor: Compose, img_batch: List[Image.Image] | torch.Tensor) -> List[Dict[str, float]]:
"""
Process a batch of images and return prediction scores for each NSFW category
"""
idx_to_label = {0: 'neutral', 1: 'low', 2: 'medium', 3: 'high'}

# Prepare batch
inputs = torch.stack([processor(img) for img in img_batch])
output = []
with torch.inference_mode():
logits = model(inputs).logits
batch_probs = F.log_softmax(logits, dim=-1)
batch_probs = torch.exp(batch_probs).cpu()

for i in r