|
|
--- |
|
|
license: apache-2.0 |
|
|
language: |
|
|
- en |
|
|
pipeline_tag: image-classification |
|
|
library_name: transformers |
|
|
tags: |
|
|
- text-generation-inference |
|
|
- siglip2 |
|
|
- image-filter |
|
|
- safe-image-moderation |
|
|
- adult-content-filter |
|
|
- content-safety |
|
|
- anime-detection |
|
|
- ai-safety |
|
|
base_model: |
|
|
- prithivMLmods/Image-Guard-ckpt-3312 |
|
|
--- |
|
|
|
|
|
 |
|
|
|
|
|
# **Image-Guard-2.0-Post0.1** |
|
|
|
|
|
> **Image-Guard-2.0-Post0.1** is a **multiclass image safety classification model** fine-tuned from **google/siglip2-base-patch16-224**. |
|
|
> It classifies images into multiple safety-related categories using the **SiglipForImageClassification** architecture. |
|
|
|
|
|
> [!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) |
|
|
|
|
|
|
|
|
```py |
|
|
Classification report: |
|
|
precision recall f1-score support |
|
|
|
|
|
Anime-SFW 0.8906 0.8766 0.8835 5600 |
|
|
Hentai 0.9081 0.8892 0.8986 4180 |
|
|
Normal-SFW 0.9010 0.8784 0.8896 5503 |
|
|
Pornography 0.9489 0.9448 0.9469 5600 |
|
|
Enticing or Sensual 0.8900 0.9436 0.9160 5600 |
|
|
|
|
|
accuracy 0.9076 26483 |
|
|
macro avg 0.9077 0.9065 0.9069 26483 |
|
|
weighted avg 0.9077 0.9076 0.9074 26483 |
|
|
``` |
|
|
|
|
|
 |
|
|
|
|
|
--- |
|
|
|
|
|
## **Label Space: 5 Classes** |
|
|
|
|
|
| Class ID | Label | Description | |
|
|
| -------- | ------------------- | ------------------------------------------------------------------------- | |
|
|
| **0** | Anime-SFW | Safe-for-work anime-style images. | |
|
|
| **1** | Hentai | Explicit or adult anime content. | |
|
|
| **2** | Normal-SFW | Realistic or photographic images that are safe for work. | |
|
|
| **3** | Pornography | Explicit adult content involving nudity or sexual acts. | |
|
|
| **4** | Enticing or Sensual | Suggestive imagery that is not explicit but intended to evoke sensuality. | |
|
|
|
|
|
--- |
|
|
|
|
|
> This model is experimental and may or may not be considered for actual use. |
|
|
|
|
|
## **Install Dependencies** |
|
|
|
|
|
```bash |
|
|
pip install -q transformers torch 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/Image-Guard-2.0-Post0.1" |
|
|
model = SiglipForImageClassification.from_pretrained(model_name) |
|
|
processor = AutoImageProcessor.from_pretrained(model_name) |
|
|
|
|
|
# Label mapping |
|
|
id2label = { |
|
|
"0": "Anime-SFW", |
|
|
"1": "Hentai", |
|
|
"2": "Normal-SFW", |
|
|
"3": "Pornography", |
|
|
"4": "Enticing or Sensual" |
|
|
} |
|
|
|
|
|
def classify_image_safety(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_safety, |
|
|
inputs=gr.Image(type="numpy"), |
|
|
outputs=gr.Label(num_top_classes=5, label="Image Safety Classification"), |
|
|
title="Image-Guard-2.0-Post0.1", |
|
|
description="Upload an image to classify it into one of five safety categories: Anime-SFW, Hentai, Normal-SFW, Pornography, or Enticing/Sensual." |
|
|
) |
|
|
|
|
|
if __name__ == "__main__": |
|
|
iface.launch() |
|
|
``` |
|
|
|
|
|
--- |
|
|
|
|
|
## **Intended Use** |
|
|
|
|
|
**Image-Guard-2.0-Post0.1** is designed for: |
|
|
|
|
|
* **Content Moderation** – Automatically identify and filter sensitive or NSFW imagery. |
|
|
* **Dataset Curation** – Separate clean and explicit data for research and training. |
|
|
* **Platform Safety** – Support compliance for social, educational, and media-sharing platforms. |
|
|
* **AI Model Input Filtering** – Prevent unsafe data from entering multimodal or generative pipelines. |
|
|
|
|
|
## **Limitations** |
|
|
|
|
|
* The model may occasionally misclassify borderline or artistically abstract images. |
|
|
* It does not perform face recognition or identify individuals. |
|
|
* Results depend on lighting, resolution, and visual context. |
|
|
* The model does not replace human moderation for sensitive environments. |
