Ai Vs Human Image Detector

🚀 AI与人类图像分类模型

本项目是一个经过微调的图像分类模型，在60,000张AI生成图像和60,000张人类图像上进行训练。该模型能够出色地检测出由Midjourney v6.1、Flux 1.1 Pro、Stable Diffusion 3.5、GPT - 4o等前沿AI模型生成的高质量图像。

详细的训练代码可查看：blog/ai/fine - tuning - siglip2

🚀 快速开始

安装依赖

pip install -q transformers torch Pillow accelerate

代码示例

import torch
from PIL import Image as PILImage
from transformers import AutoImageProcessor, SiglipForImageClassification

MODEL_IDENTIFIER = r"Ateeqq/ai-vs-human-image-detector"

# Device: Use GPU if available, otherwise CPU
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print(f"Using device: {device}")

# Load Model and Processor
try:
    print(f"Loading processor from: {MODEL_IDENTIFIER}")
    processor = AutoImageProcessor.from_pretrained(MODEL_IDENTIFIER)

    print(f"Loading model from: {MODEL_IDENTIFIER}")
    model = SiglipForImageClassification.from_pretrained(MODEL_IDENTIFIER)
    model.to(device)
    model.eval()
    print("Model and processor loaded successfully.")

except Exception as e:
    print(f"Error loading model or processor: {e}")
    exit()

# Load and Preprocess the Image

IMAGE_PATH = r"/content/images.jpg" 
try:
    print(f"Loading image: {IMAGE_PATH}")
    image = PILImage.open(IMAGE_PATH).convert("RGB")
except FileNotFoundError:
    print(f"Error: Image file not found at {IMAGE_PATH}")
    exit()
except Exception as e:
    print(f"Error opening image: {e}")
    exit()

print("Preprocessing image...")
# Use the processor to prepare the image for the model
inputs = processor(images=image, return_tensors="pt").to(device)

# Perform Inference
print("Running inference...")
with torch.no_grad(): # Disable gradient calculations for inference
    outputs = model(**inputs)
    logits = outputs.logits

# Interpret the Results
# Get the index of the highest logit score -> this is the predicted class ID
predicted_class_idx = logits.argmax(-1).item()

# Use the model's config to map the ID back to the label string ('ai' or 'hum')
predicted_label = model.config.id2label[predicted_class_idx]

# Optional: Get probabilities using softmax
probabilities = torch.softmax(logits, dim=-1)
predicted_prob = probabilities[0, predicted_class_idx].item()

print("-" * 30)
print(f"Image: {IMAGE_PATH}")
print(f"Predicted Label: {predicted_label}")
print(f"Confidence Score: {predicted_prob:.4f}")
print("-" * 30)

# You can also print the scores for all classes:
print("Scores per class:")
for i, label in model.config.id2label.items():
    print(f"  - {label}: {probabilities[0, i].item():.4f}")

输出示例

Using device: cpu
Model and processor loaded successfully.
Loading image: /content/images.jpg
Preprocessing image...
Running inference...
------------------------------
Image: /content/images.jpg
Predicted Label: ai
Confidence Score: 0.9996
------------------------------
Scores per class:
  - ai: 0.9996
  - hum: 0.0004

✨ 主要特性

• 强大的检测能力：能够有效识别多种前沿AI模型生成的高质量图像。
• 高准确率：在测试集上取得了高达0.9923的准确率。

📚 详细文档

评估指标

🏋️‍♂️ 训练指标

• Epoch：5.0
• 总浮点运算次数（Total FLOPs）：51,652,280,821 GF
• 训练损失（Train Loss）：0.0799
• 训练运行时间（Train Runtime）：2:39:49.46
• 训练样本每秒处理数（Train Samples/Sec）：69.053
• 训练步骤每秒处理数（Train Steps/Sec）：4.316

📊 评估指标（微调模型在测试集上）

• Epoch：5.0
• 评估准确率（Eval Accuracy）：0.9923
• 评估损失（Eval Loss）：0.0551
• 评估运行时间（Eval Runtime）：0:02:35.78
• 评估样本每秒处理数（Eval Samples/Sec）：212.533
• 评估步骤每秒处理数（Eval Steps/Sec）：6.644

🔦 预测指标（在测试集上）

{
  "test_loss": 0.05508904904127121,
  "test_accuracy": 0.9923283699296264,
  "test_runtime": 167.1844,
  "test_samples_per_second": 198.039,
  "test_steps_per_second": 6.191
}

• 最终测试准确率：0.9923
• 最终测试F1分数（Macro）：0.9923
• 最终测试F1分数（Weighted）：0.9923

📄 许可证

本项目采用Apache - 2.0许可证。