---
license: apache-2.0
base_model:
- google/siglip2-base-patch16-224
library_name: transformers
tags:
- SigLIP2
- ImageShield
- 90M
- Guardrail
language:
- en
pipeline_tag: image-classification
datasets:
- prithivMLmods/ImageShield-Guardrail-80K
- prithivMLmods/ImageShield-Guardrail-Realism-60K
---

![1](https://cdn-uploads.huggingface.co/production/uploads/65bb837dbfb878f46c77de4c/PxMC3xQE3LdbpfWzv4MBD.png)

# **ImageShield-SUPER-90M**

> **ImageShield-SUPER-90M** is a vision-language image classification model based on **google/siglip2-base-patch16-224**, trained on **100K samples from the ImageShield-Guardrail Safe and Unsafe Images dataset**. Built on the **SiglipForImageClassification** architecture, the model is designed to classify visual content as **Safe** or **Unsafe** for content moderation and media filtering.

> [!IMPORTANT]
> This model is experimental. Expert multimodal models are available here: [ImageShield Multimodal SFT Collection](https://huggingface.co/collections/prithivMLmods/imageshield-multimodal-sft).

> [!note]
> *SigLIP 2: Multilingual Vision-Language Encoders with Improved Semantic Understanding, Localization, and Dense Features*
> [https://arxiv.org/pdf/2502.14786](https://arxiv.org/pdf/2502.14786)

## **Label Space: 2 Classes**

The model classifies each image into one of the following content categories:

```text
Class 0: "Safe"
Class 1: "Unsafe"
```

## **Install Dependencies**

```bash
pip install transformers torch torchvision pillow gradio
```

## **Inference Code**

```python
import gradio as gr
from transformers import AutoImageProcessor, SiglipForImageClassification
from PIL import Image
import torch

# Load model and processor
model_name = "prithivMLmods/ImageShield-SUPER-90M"
model = SiglipForImageClassification.from_pretrained(model_name)
processor = AutoImageProcessor.from_pretrained(model_name)

# ID to Label mapping
id2label = {
    "0": "Safe",
    "1": "Unsafe"
}

def classify_image(image):
    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()

    prediction = {
        id2label[str(i)]: round(probs[i], 3)
        for i in range(len(probs))
    }

    return prediction

# Gradio Interface
iface = gr.Interface(
    fn=classify_image,
    inputs=gr.Image(type="numpy"),
    outputs=gr.Label(
        num_top_classes=2,
        label="Predicted Content Type"
    ),
    title="ImageShield-SUPER-90M",
    description="Classifies images as Safe or Unsafe."
)

if __name__ == "__main__":
    iface.launch()
```

## **Intended Use**

This model is intended for applications such as:

* **Content Moderation:** Identify unsafe visual content.
* **Parental Controls:** Support AI-based media filtering.
* **Dataset Preprocessing:** Categorize and filter safe and unsafe images.
* **Online Platforms:** Assist with content safety and upload moderation.
* **AI Image Applications:** Provide an additional safety layer for image generation and editing workflows.

## **Classification Report**

### Training vs Evaluation Loss / Accuracy

![Training vs Evaluation Loss and Accuracy](assets/training_eval_graph.png)

### Precision / Recall / F1-score per Class

![Per-Class Precision, Recall, and F1-score](assets/classification_report_bar.png)

### Confusion Matrix

![Confusion Matrix](assets/confusion_matrix.png)

### Test Set Class Distribution

![Test Set Class Distribution](assets/class_distribution_pie.png)

### Overall Prediction Accuracy

![Overall Prediction Accuracy](assets/prediction_accuracy_pie.png)

### Misalignment Distribution by True Class

![Misalignment Distribution by True Class](assets/misalignment_distribution_pie.png)

## **Acknowledgements**

* **[Transformers](https://huggingface.co/docs/transformers/en/index)**: Transformers provides state-of-the-art machine learning models for text, computer vision, audio, video, and multimodal tasks, supporting both inference and training.

* **[SigLIP 2](https://huggingface.co/papers/2502.14786)**: Multilingual vision-language encoders with improved semantic understanding, localization, and dense feature representations.