Guard-Against-Unsafe-Content2-Siglip2

Guard-Against-Unsafe-Content2-Siglip2 is an image classification vision-language encoder model fine-tuned from google/siglip2-base-patch16-224 for a binary classification task. It is designed to classify images as either "normal" or "nsfw" using the SiglipForImageClassification architecture.

Experimental: This NSFW filter is an experimental model. Since I haven't found a better dataset to improve it, its performance may be inconsistent in some cases. I am currently looking for a better open dataset to enhance its effectiveness in a multi-label classification problem.

Classification Report:
              precision    recall  f1-score   support

      normal     0.9975    0.9988    0.9981      4000
        nsfw     0.9992    0.9983    0.9987      6000

    accuracy                         0.9985     10000
   macro avg     0.9983    0.9985    0.9984     10000
weighted avg     0.9985    0.9985    0.9985     10000

The model categorizes images into 2 classes:

• Class 0: "normal"
• Class 1: "nsfw"

Run with Transformers🤗

!pip install -q transformers torch pillow gradio

import gradio as gr
from transformers import AutoImageProcessor
from transformers import SiglipForImageClassification
from transformers.image_utils import load_image
from PIL import Image
import torch

# Load model and processor
model_name = "prithivMLmods/Guard-Against-Unsafe-Content2-Siglip2"
model = SiglipForImageClassification.from_pretrained(model_name)
processor = AutoImageProcessor.from_pretrained(model_name)

def nsfw_classification(image):
    """Predicts whether an image is NSFW or normal."""
    image = Image.fromarray(image).convert("RGB")
    inputs = processor(images=image, return_tensors="pt")
    
    with torch.no_grad():
        outputs = model(**inputs)
        logits = outputs.logits
        probs = torch.nn.functional.softmax(logits, dim=1).squeeze().tolist()
    
    labels = {
        "0": "normal", 
        "1": "nsfw"
    }
    predictions = {labels[str(i)]: round(probs[i], 3) for i in range(len(probs))}
    
    return predictions

# Create Gradio interface
iface = gr.Interface(
    fn=nsfw_classification,
    inputs=gr.Image(type="numpy"),
    outputs=gr.Label(label="Prediction Scores"),
    title="NSFW Image Classification",
    description="Upload an image to classify whether it is normal or NSFW."
)

# Launch the app
if __name__ == "__main__":
    iface.launch()

Intended Use:

The Guard-Against-Unsafe-Content2-Siglip2 model is designed to classify images as either "normal" or "nsfw". Potential use cases include:

• Content Moderation: Automatically filtering explicit images on social media, forums, and messaging apps.
• Parental Control: Ensuring children are not exposed to inappropriate content.
• Image Search Filtering: Improving search engine safety by removing NSFW images.
• Automated Image Analysis: Enhancing AI-driven applications that require content safety monitoring.