Property	Value
Base model	`facebook/deit-small-patch16-224`
Task	Binary image classification
Labels	`sfw` (0), `nsfw` (1)
Input size	224x224
Parameters	7.09M total, 21.6M fine-tuned (32.8%)

Performance

Metric	Score
Accuracy	98.00%
Precision	98.76%
Recall	97.15%
F1	97.95%

Training

Dataset: 10,000 images — 5,000 SFW (coding/cs related images) from CodeSCAN and handcrafted screenshots + 5,000 NSFW (1,000 each from drawings, hentai, neutral, porn, sexy) from deepghs/nsfw_detect
Split: 80% train / 10% val / 10% test
Strategy: Froze early layers, fine-tuned last 4 transformer blocks + classifier head
Hyperparameters: lr=2e-5, batch_size=32, cosine schedule, 10 epochs with early stopping (patience=3)
Augmentation: Random crop, horizontal flip, rotation, color jitter

Usage

ONNX Runtime

python

from optimum.onnxruntime import ORTModelForImageClassification
from transformers import DeiTImageProcessor
import torch
import requests
from PIL import Image

model_name = "navodPeiris/deit-nsfw"

# Load ONNX model
model = ORTModelForImageClassification.from_pretrained(model_name, subfolder="onnx")
processor = DeiTImageProcessor.from_pretrained(model_name)

# Image URL
url = "https://www.cyberschoolers.com/wp-content/uploads/ztm-web-developer-course-curriculum.webp"

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

# Preprocess
inputs = processor(images=image, return_tensors="pt")

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

logits = outputs.logits

# Convert to probabilities
probs = torch.softmax(logits, dim=-1)

# Predicted class
pred_id = torch.argmax(probs, dim=-1).item()
label = model.config.id2label[pred_id]
confidence = probs[0][pred_id].item()

print(label, f"{confidence}")

Transformers.js

javascript

import { pipeline } from "@huggingface/transformers";
const classifier = await pipeline(
  "image-classification",
  "navodPeiris/deit-nsfw",
);
const result = await classifier("image.jpg");

Limitations

Trained primarily on code screenshots, CS related images, CS website screenshots as SFW examples — may not generalize well to all SFW content types
Binary classification only — does not distinguish between NSFW subcategories
The "neutral" category from the NSFW dataset was included in the NSFW class to filter even neutral images while strictly only allowing codings/cs related content