vit-nsfw-stable-diffusion

This model is a fine-tuned version of vit-base-patch16-384 on ~1.7M generated images from the AdamCodd/Civitai-8m-prompts dataset, balanced between NSFW/SFW labels. It achieves the following results on the evaluation set:

• Loss: 0.1592
• Accuracy: 0.9349

Unlike AdamCodd/vit-base-nsfw-detector model, this one was exclusively trained on generated images from stable diffusion.

The license for this model is cc-by-nc-nd-4.0. For commercial use rights, please contact me (adamcoddml@gmail.com).

Model description

The Vision Transformer (ViT) is a transformer encoder model (BERT-like) pretrained on a large collection of images in a supervised fashion, namely ImageNet-21k, at a resolution of 224x224 pixels. Next, the model was fine-tuned on ImageNet (also referred to as ILSVRC2012), a dataset comprising 1 million images and 1,000 classes, at a higher resolution of 384x384.

Intended uses & limitations

Usage for a local image:

from transformers import pipeline
from PIL import Image

img = Image.open("<path_to_image_file>")
predict = pipeline('image-classification', model='AdamCodd/vit-nsfw-stable-diffusion')
predict(img)

Usage for a distant image:

from transformers import ViTImageProcessor, AutoModelForImageClassification
from PIL import Image
import requests

url = 'https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/ba55f276-5aa5-446f-c59a-8fff4d209100/width=512/ba55f276-5aa5-446f-c59a-8fff4d209100.jpeg'
image = Image.open(requests.get(url, stream=True).raw)
processor = ViTImageProcessor.from_pretrained('AdamCodd/vit-nsfw-stable-diffusion')
model = AutoModelForImageClassification.from_pretrained('AdamCodd/vit-nsfw-stable-diffusion')
inputs = processor(images=image, return_tensors="pt")
outputs = model(**inputs)
logits = outputs.logits

predicted_class_idx = logits.argmax(-1).item()
print("Predicted class:", model.config.id2label[predicted_class_idx])
# Predicted class: sfw

Usage with Transformers.js (Vanilla JS):

/* Instructions:
* - Place this script in an HTML file using the <script type="module"> tag.
* - Ensure the HTML file is served over a local or remote server (e.g., using Python's http.server, Node.js server, or similar).
* - Replace 'https://example.com/path/to/image.jpg' in the classifyImage function call with the URL of the image you want to classify.
*
* Example of how to include this script in HTML:
* <script type="module" src="path/to/this_script.js"></script>
*
* This setup ensures that the script can use imports and perform network requests without CORS issues.
*/
import { pipeline, env } from 'https://cdn.jsdelivr.net/npm/@xenova/transformers@2.17.1';

// Since we will download the model from HuggingFace Hub, we can skip the local model check
env.allowLocalModels = false;

// Load the image classification model
const classifier = await pipeline('image-classification', 'AdamCodd/vit-nsfw-stable-diffusion');

// Function to fetch and classify an image from a URL
async function classifyImage(url) {
  try {
    const response = await fetch(url);
    if (!response.ok) throw new Error('Failed to load image');

    const blob = await response.blob();
    const image = new Image();
    const imagePromise = new Promise((resolve, reject) => {
      image.onload = () => resolve(image);
      image.onerror = reject;
      image.src = URL.createObjectURL(blob);
    });

    const img = await imagePromise; // Ensure the image is loaded
    const classificationResults = await classifier([img.src]); // Classify the image
    console.log('Predicted class: ', classificationResults[0].label);
  } catch (error) {
    console.error('Error classifying image:', error);
  }
}

// Example usage
classifyImage('https://example.com/path/to/image.jpg');
// Predicted class: sfw

Since this model has been trained on generated images from stable diffusion, it won't perform as well on real pictures (in that case just use my other ViT model).

It performs very well on generated images and would pair well with the following AdamCodd/distilroberta-nsfw-prompt-stable-diffusion model to filter both prompts and images, ensuring safe results.

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 3e-05
• train_batch_size: 32
• eval_batch_size: 64
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• num_epochs: 2

Training results

• Validation Loss: 0.1592
• Accuracy: 0.9349
• F1: 0.9350
• AUC: 0.9847
• Precision: 0.9335
• Recall: 0.9366

Confusion matrix (eval):

[[78666 5644]

[5355 79173]]

Framework versions

• Transformers 4.36.2
• Evaluate 0.4.1

If you want to support me, you can here.

Citation and Acknowledgments

I would like to express my sincere gratitude to Prodia.com for generously providing the GPU resources, specifically the RTX 4090, that made the training of this model possible.