This model is a Swin Transformer-based classifier designed to distinguish between AI-generated and human-created images, specifically focusing on outputs from Midjourney V6 and Stable Diffusion XL (SDXL). It has been trained on a curated dataset of AI-generated images.

Developed by: Deepankar Sharma
Model type: Image Classification (Swin Transformer)
Finetuned from model: SwinForImageClassification

Model Sources

Repository: Hugging Face Model Repository

Uses

Direct Use

This model can be used for detecting AI-generated images from Midjourney V6 and SDXL. It is useful for content moderation, fact-checking, and detecting synthetic media.

Out-of-Scope Use

The model is not designed for detecting AI-generated images from all generative models.
It may not perform well on heavily edited AI-generated images or images mixed with human elements.
It is not intended for forensic-level deepfake detection.

Bias, Risks, and Limitations

This model is trained specifically on Midjourney V6 and Stable Diffusion XL datasets. It may not generalize well to images generated by other AI models. Additionally, biases in the dataset could lead to false positives (flagging real images as AI-generated) or false negatives (failing to detect AI-generated content).

Recommendations

Users should verify results with additional tools and not solely rely on this model for high-stakes decisions. Model performance should be tested on domain-specific datasets before deployment.

How to Get Started with the Model

You can use this model with the 🤗 Transformers library:

python

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

# Load model and feature extractor
model_name = "ideepankarsharma2003/AI_ImageClassification_MidjourneyV6_SDXL"
model = AutoModelForImageClassification.from_pretrained(model_name)
feature_extractor = AutoFeatureExtractor.from_pretrained(model_name)

# Load and preprocess image
image = Image.open("path_to_image.jpg")
inputs = feature_extractor(images=image, return_tensors="pt")

# Perform inference
with torch.no_grad():
    outputs = model(**inputs)
    logits = outputs.logits
    predicted_label = logits.argmax(-1).item()

# Label Mapping
id2label = {0: "ai_gen", 1: "human"}
print("Predicted label:", id2label[predicted_label])

Training Details

Training Data

The model was trained on the following datasets:

Training Procedure

Image Size: 224x224
Patch Size: 4
Embedding Dimension: 128
Layers: 4
Attention Heads per Stage: [4, 8, 16, 32]
Dropout Rates:
- Attention: 0.0
- Hidden: 0.0
- Drop Path: 0.1
Activation Function: GeLU
Optimizer: AdamW
Learning Rate Scheduler: Cosine Annealing
Precision: float32
Training Steps: 3414

Evaluation

Testing Data, Factors & Metrics

Testing Data

The model was evaluated on a separate validation split from the training datasets.

Metrics

Accuracy
Precision & Recall
F1 Score

Summary

The model effectively distinguishes between AI-generated and human-created images, but its performance may be affected by dataset biases and out-of-distribution examples.

Citation

If you use this model, please cite:

bibtex

@misc{ai_image_classification,
  author = {Deepankar Sharma},
  title = {AI Image Classification - Midjourney V6 & SDXL},
  year = {2024},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/ideepankarsharma2003/AI_ImageClassification_MidjourneyV6_SDXL}}
}

Model Card Authors

Author: Deepankar Sharma