Image Safety Classifier

NOTE: Like all models, this one can make mistakes. NSFW content can be subjective and contextual, this model is intended to help identify this content, use at your own risk.

OwenElliott/image-safety-classifier-s is a lightweight image classification model designed to categorise images as either NSFW, NSFL, or SFW. NSFW images contain pornographic or highly suggestive content, NSFL images contain gore, and SFW images are everything else. The small size of this model and the SwiftFormer architecture make this model suitable for edge deployment or latency critical applications.

The model is trained on a proprietary dataset of ~320,000 images scraped from the web. These images include a diverse range of content including real photos, drawings, Rule 34, screenshots, AI generated images, memes, and more. The NSFW and SFW classes were checked with Marqo/nsfw-image-detection-384 and any images which were confidently predicted to be in the wrong class were manually reviewed.

Model Family

This model is part of a lineup of image safety classification models in 4 different sizes. All models are finetunes of SwiftFormer models.

Evaluation

This model outperforms existing NSFW detectors on our dataset, here we provide an evaluation against Marqo/nsfw-image-detection-384 and Falconsai/nsfw_image_detection.

*Note that Marqo/nsfw-image-detection-384 and Falconsai/nsfw_image_detection don't have an explicit NSFL class, for this evaluation the NSFW class from these model is treated as a correct classification for an NSFL image.

Model Usage

Image Classification with timm

pip install timm

from urllib.request import urlopen
from PIL import Image
import timm
import torch

img = Image.open(urlopen(
    'https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/beignets-task-guide.png'
))

model = timm.create_model("OwenElliott/image-safety-classifier-s", pretrained=True)
model = model.eval()

data_config = timm.data.resolve_model_data_config(model)
transforms = timm.data.create_transform(**data_config, is_training=False)

with torch.no_grad():
    output = model(transforms(img).unsqueeze(0)).softmax(dim=-1).cpu()

class_names = model.pretrained_cfg["label_names"]
print("Probabilities:", output[0])
print("Class:", class_names[output[0].argmax()])

ONNX Inference

ONNX models are available with preprocessing baked in — normalization and softmax are part of the graph, so you just need to resize your image and pass pixel values in 0-255 range. Both fp32 and fp16 variants are provided.

pip install onnxruntime pillow numpy
# For GPU: pip install onnxruntime-gpu

from urllib.request import urlopen
from PIL import Image
import numpy as np
import onnxruntime as ort

img = Image.open(urlopen(
    'https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/beignets-task-guide.png'
))

# The model accepts 224x224 images
img = img.convert("RGB").resize((224, 224), Image.BILINEAR)

# Convert to [1, 3, 224, 224] float32 in 0-255
pixels = np.array(img, dtype=np.float32).transpose(2, 0, 1)[np.newaxis]

# Run inference (use fp16 variant for GPU acceleration)
# image colour channel normalisation is baked into the ONNX models
from huggingface_hub import hf_hub_download
model_path = hf_hub_download("OwenElliott/image-safety-classifier-s", "image-safety-classifier-s.onnx")
sess = ort.InferenceSession(model_path)
probs = sess.run(None, {{"image": pixels}})[0][0]

class_names = ["NSFL", "NSFW", "SFW"]
print("Probabilities:", dict(zip(class_names, probs)))
print("Class:", class_names[np.argmax(probs)])

Core ML conversion

The repository includes a conversion script that creates an iOS/macOS Core ML package from model.safetensors. The resulting model has an RGB image input, performs the same ImageNet normalization and softmax as the ONNX model, and returns classLabel plus classLabel_probs for NSFL, NSFW, and SFW.

pip install coremltools timm safetensors torch
python convert_to_coreml.py --output ImageSafetyClassifier.mlpackage

Use --fp16 for a smaller/faster package on Apple hardware, or select a different minimum OS with --deployment-target ios17 or --deployment-target macos14. The caller should resize and center-crop the image to 224x224; the Core ML image input handles the 0..255 to 0..1 conversion.

In Swift, the generated package can be loaded with VNCoreMLModel or directly with Core ML. For Vision, use VNCoreMLRequest with a CVPixelBuffer; for direct Core ML prediction, provide the image feature as a 224x224 RGB image.

Notes on the Dataset

The dataset is primarily NSFW and SFW examples, the NSFL class is very underrepresented, collecting NSFL data at scale has been challenging. This is something I would like to improve in a future version, for the moment this model significatly outperforms other models on NSFL classification however there is definitely room for improvement.

Citations

@InProceedings{Shaker_2023_ICCV,
    author    = {Shaker, Abdelrahman and Maaz, Muhammad and Rasheed, Hanoona and Khan, Salman and Yang, Ming-Hsuan and Khan, Fahad Shahbaz},
    title     = {SwiftFormer: Efficient Additive Attention for Transformer-based Real-time Mobile Vision Applications},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
    year      = {2023},
}

@misc{rw2019timm,
  author = {Ross Wightman},
  title = {PyTorch Image Models},
  year = {2019},
  publisher = {GitHub},
  journal = {GitHub repository},
  doi = {10.5281/zenodo.4414861},
  howpublished = {\url{https://github.com/huggingface/pytorch-image-models}}
}