A Vision Transformer (ViT) for binary NSFW/SFW image classification, specifically trained to handle fetish content and a wide variety of image styles including artwork, illustrations, and AI-generated imagery.

Model Description

This model is a fine-tuned version of AdamCodd/vit-base-nsfw-detector, further trained on a curated dataset to improve detection of fetish-specific NSFW content and to generalize across diverse image types.

What's Different From the Base Model

The base model (AdamCodd/vit-base-nsfw-detector) was primarily trained on photographic content. This fine-tune extends its capabilities to better handle:

Fetish-specific content — BDSM, latex, leather, and related imagery
Photographs — Real-world photos of various styles
AI-generated images — Content from Stable Diffusion, Midjourney, DALL-E, etc.
Digital illustrations — Anime, manga, cartoon styles
Traditional artwork — Drawings, paintings, sketches
3D renders — CGI and rendered artwork

This makes the model more robust for content moderation systems that need to handle mixed media sources.

Model Details

Architecture: Vision Transformer (ViT-base, patch size 16)
Input resolution: 384×384 RGB images
Output: Binary classification (sfw, nsfw)
Base model: AdamCodd/vit-base-nsfw-detector

Performance

Metric	Validation	Test (Held-Out)
Accuracy	93.95%	92.44%
F1 Score	93.95%	92.44%
Precision	93.97%	92.45%
Recall	93.95%	92.44%

Training Details

Training samples: 1,862
Validation samples: 397
Test samples: 397 (held-out)
Epochs: 10
Final training loss: 0.031

Dataset Composition

Label	Total	Train	Val	Test
NSFW	1,351	947	202	202
SFW	1,305	915	195	195

The dataset includes a balanced mix of photographic content, AI-generated images, digital illustrations, drawings, and 3D renders across both classes.

Usage

python

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

# Load model and processor
model_name = "electrohead/vit-fetish-nsfw-detector"
processor = AutoImageProcessor.from_pretrained(model_name)
model = AutoModelForImageClassification.from_pretrained(model_name)

# Load and process image
image = Image.open("your_image.jpg")
inputs = processor(images=image, return_tensors="pt")

# Get prediction
with torch.no_grad():
    outputs = model(**inputs)
    predicted_class = outputs.logits.argmax(-1).item()

label = model.config.id2label[predicted_class]
print(f"Prediction: {label}")

# Get probabilities
probs = torch.nn.functional.softmax(outputs.logits, dim=-1)
print(f"SFW: {probs[0][0]:.2%}, NSFW: {probs[0][1]:.2%}")

Or using the pipeline API:

python

from transformers import pipeline

classifier = pipeline("image-classification", model="electrohead/vit-fetish-nsfw-detector")
result = classifier("your_image.jpg")
print(result)

Intended Use

This model is designed for:

Content moderation systems handling diverse media types
Automated content filtering for platforms with user-generated content
Pre-screening uploads that may include art, AI images, or photographs
Research on NSFW content detection across image modalities

Limitations

Fetish focus: While the model handles general NSFW content, it was specifically tuned for fetish-related imagery. Performance on other NSFW categories may differ from the base model.
Binary classification: No severity levels or subcategories — images are classified only as sfw or nsfw.
Edge cases: Artistic nudity, medical imagery, and ambiguous content may produce inconsistent results.
Style bias: Despite training on diverse styles, novel artistic techniques may not be well-represented.

Ethical Considerations

This model is intended for legitimate content moderation purposes. Please use responsibly.

Not intended for:

Surveillance or privacy-invasive applications
Discriminatory filtering based on body types or identities
Circumventing platform policies
Any illegal purposes

Acknowledgments

Built on AdamCodd/vit-base-nsfw-detector
Fine-tuned using HuggingFace Transformers
Training managed with Label Studio ML Backend