Nayatel Vision Transformer NSFW Detector (ViT-Large)

This model is a fine-tuned Vision Transformer (ViT-Large, Patch16-224) developed by the Nayatel AI & DevOps Team to classify images into NSFW or Safe categories.

It is designed for high-recall NSFW detection, making it suitable for content moderation, filtering pipelines, and safety-critical applications.

📊 Evaluation Metrics

Evaluation was performed on a held-out validation set derived from a self-curated dataset. The model demonstrates strong generalization and robustness across diverse NSFW and safe image categories.

🧠 Model Description

• Architecture: Vision Transformer ViT-Large (Patch16, 224×224)
• Parameters: ~304 million
• Base Model: google/vit-large-patch16-224
• Pretrained on: ImageNet-21k
• Fine-tuned by: Nayatel AI & DevOps Team
• Classes: nsfw, safe
• Label Mapping:
  • 0 → nsfw
  • 1 → safe
• Output: Softmax probabilities over two classes

ViT splits an image into fixed-size patches and processes them using multi-head self-attention across 24 transformer layers, enabling strong global context understanding.

📁 Training Details

• Epochs: 5
• Batch Size: 32
• Learning Rate: 3e-5
• Optimizer: AdamW
• Loss Function: Weighted Cross-Entropy (to handle class imbalance)
• Evaluation Strategy: Per epoch
• Precision: FP16 (mixed precision)
• Frameworks: PyTorch, Hugging Face Transformers, Datasets
• Hardware: NVIDIA L4 (24 GB VRAM)

🧪 Dataset

The model was trained on a self-curated, large-scale dataset compiled from multiple sources and internal filtering pipelines.

Dataset Split

Dataset Characteristics

• Binary labels: safe / nsfw
• Content: Includes explicit, borderline, artistic, and real-world NSFW content.
• Variety: Large variety of safe images (people, objects, environments, graphics).
• Diversity: Diverse ethnicities, lighting conditions, poses, and content styles.
• Quality: Curated to reduce noise and label ambiguity.

✅ Intended Uses

• Content moderation for web and mobile applications.
• Upload filtering and pre-screening pipelines.
• Dataset sanitization for AI/ML workflows.
• ISP-level and enterprise safety enforcement.
• Parental control and compliance systems.

⚠️ Limitations

• Borderline cases (e.g., swimwear, medical imagery, artistic nudity) may be ambiguous.
• Cultural definitions of NSFW content may vary.
• Binary classification only (no fine-grained categories).
• Recommended to use threshold tuning and optional human review in sensitive applications.

🚀 How to Use

Using the Hugging Face Pipeline

from transformers import pipeline
from PIL import Image

predict = pipeline(
    "image-classification",
    model="theusamaaslam/vit-Large-nsfw-detector",
    device=0  # GPU
)

img = Image.open("<path_to_image_file>").convert("RGB")
result = predict(img)
print(result)

Usage for a distant image:

from transformers import AutoImageProcessor, AutoModelForImageClassification
from PIL import Image
import requests
import torch

# URL of the image to test
url = 'http://images.cocodataset.org/val2017/000000039769.jpg'

# Load image from URL
image = Image.open(requests.get(url, stream=True).raw).convert("RGB")

# Load Processor and Model from your Hub repo
repo_id = "theusamaaslam/vit-Large-nsfw-detector"
processor = AutoImageProcessor.from_pretrained(repo_id)
model = AutoModelForImageClassification.from_pretrained(repo_id)

# Process image and inference
inputs = processor(images=image, return_tensors="pt")

with torch.no_grad():
    outputs = model(**inputs)

# Get probabilities and predicted label
logits = outputs.logits
probs = logits.softmax(dim=1)
predicted_class_idx = logits.argmax(-1).item()
predicted_label = model.config.id2label[predicted_class_idx]

print(f"Predicted class: {predicted_label}")
print(f"Probabilities: {probs}")

📬 Contact

Author: @theusamaaslam
Email: [usamaaslam081@gmail.com]
Organization: Nayatel (Pvt Ltd) AI & DevOps Team