import gradio as gr
import torch
from transformers import AutoFeatureExtractor, AutoModelForImageClassification, pipeline
import os
import zipfile
import shutil
import matplotlib.pyplot as plt
from sklearn.metrics import accuracy_score, roc_auc_score, confusion_matrix, classification_report, roc_curve, auc
from tqdm import tqdm
from PIL import Image
import uuid
import tempfile
import pandas as pd
from numpy import exp
import numpy as np
from sklearn.metrics import ConfusionMatrixDisplay
import urllib.request
# Define models
models = [
"umm-maybe/AI-image-detector",
"Organika/sdxl-detector",
"cmckinle/sdxl-flux-detector",
]
pipe0 = pipeline("image-classification", f"{models[0]}")
pipe1 = pipeline("image-classification", f"{models[1]}")
pipe2 = pipeline("image-classification", f"{models[2]}")
fin_sum = []
uid = uuid.uuid4()
# Softmax function
def softmax(vector):
e = exp(vector - vector.max()) # for numerical stability
return e / e.sum()
# Single image classification functions
def image_classifier0(image):
labels = ["AI", "Real"]
outputs = pipe0(image)
results = {}
for idx, result in enumerate(outputs):
results[labels[idx]] = float(outputs[idx]['score']) # Convert to float
fin_sum.append(results)
return results
def image_classifier1(image):
labels = ["AI", "Real"]
outputs = pipe1(image)
results = {}
for idx, result in enumerate(outputs):
results[labels[idx]] = float(outputs[idx]['score']) # Convert to float
fin_sum.append(results)
return results
def image_classifier2(image):
labels = ["AI", "Real"]
outputs = pipe2(image)
results = {}
for idx, result in enumerate(outputs):
results[labels[idx]] = float(outputs[idx]['score']) # Convert to float
fin_sum.append(results)
return results
def aiornot0(image):
labels = ["AI", "Real"]
mod = models[0]
feature_extractor0 = AutoFeatureExtractor.from_pretrained(mod)
model0 = AutoModelForImageClassification.from_pretrained(mod)
input = feature_extractor0(image, return_tensors="pt")
with torch.no_grad():
outputs = model0(**input)
logits = outputs.logits
probability = softmax(logits) # Apply softmax on logits
px = pd.DataFrame(probability.numpy())
prediction = logits.argmax(-1).item()
label = labels[prediction]
html_out = f"""
This image is likely: {label}
Probabilities:
Real: {float(px[1][0]):.4f}
AI: {float(px[0][0]):.4f}"""
results = {
"Real": float(px[1][0]),
"AI": float(px[0][0])
}
fin_sum.append(results)
return gr.HTML.update(html_out), results
def aiornot1(image):
labels = ["AI", "Real"]
mod = models[1]
feature_extractor1 = AutoFeatureExtractor.from_pretrained(mod)
model1 = AutoModelForImageClassification.from_pretrained(mod)
input = feature_extractor1(image, return_tensors="pt")
with torch.no_grad():
outputs = model1(**input)
logits = outputs.logits
probability = softmax(logits) # Apply softmax on logits
px = pd.DataFrame(probability.numpy())
prediction = logits.argmax(-1).item()
label = labels[prediction]
html_out = f"""
This image is likely: {label}
Probabilities:
Real: {float(px[1][0]):.4f}
AI: {float(px[0][0]):.4f}"""
results = {
"Real": float(px[1][0]),
"AI": float(px[0][0])
}
fin_sum.append(results)
return gr.HTML.update(html_out), results
def aiornot2(image):
labels = ["AI", "Real"]
mod = models[2]
feature_extractor2 = AutoFeatureExtractor.from_pretrained(mod)
model2 = AutoModelForImageClassification.from_pretrained(mod)
input = feature_extractor2(image, return_tensors="pt")
with torch.no_grad():
outputs = model2(**input)
logits = outputs.logits
probability = softmax(logits) # Apply softmax on logits
px = pd.DataFrame(probability.numpy())
prediction = logits.argmax(-1).item()
label = labels[prediction]
html_out = f"""
This image is likely: {label}
Probabilities:
Real: {float(px[1][0]):.4f}
AI: {float(px[0][0]):.4f}"""
results = {
"Real": float(px[1][0]),
"AI": float(px[0][0])
}
fin_sum.append(results)
return gr.HTML.update(html_out), results
# Function to extract images from zip
def extract_zip(zip_file):
temp_dir = tempfile.mkdtemp() # Temporary directory
with zipfile.ZipFile(zip_file, 'r') as z:
z.extractall(temp_dir)
return temp_dir
# Function to classify images in a folder
def classify_images(image_dir, model_pipeline, model_idx):
images = []
labels = []
preds = []
for folder_name, ground_truth_label in [('real', 1), ('ai', 0)]:
folder_path = os.path.join(image_dir, folder_name)
if not os.path.exists(folder_path):
print(f"Folder not found: {folder_path}")
continue
for img_name in os.listdir(folder_path):
img_path = os.path.join(folder_path, img_name)
try:
img = Image.open(img_path).convert("RGB")
# Ensure that each image is being processed by the correct model pipeline
pred = model_pipeline(img)
pred_label = 0 if pred[0]['label'] == 'AI' else 1 # Assuming 'AI' is label 0 and 'Real' is label 1
preds.append(pred_label)
labels.append(ground_truth_label)
images.append(img_name)
except Exception as e:
print(f"Error processing image {img_name} in model {model_idx}: {e}")
print(f"Model {model_idx} processed {len(images)} images")
return labels, preds, images
# Function to generate evaluation metrics
def evaluate_model(labels, preds):
cm = confusion_matrix(labels, preds)
accuracy = accuracy_score(labels, preds)
roc_score = roc_auc_score(labels, preds)
report = classification_report(labels, preds)
fpr, tpr, _ = roc_curve(labels, preds)
roc_auc = auc(fpr, tpr)
fig, ax = plt.subplots()
disp = ConfusionMatrixDisplay(confusion_matrix=cm, display_labels=["AI", "Real"])
disp.plot(cmap=plt.cm.Blues, ax=ax)
plt.close(fig)
fig_roc, ax_roc = plt.subplots()
ax_roc.plot(fpr, tpr, color='blue', lw=2, label=f'ROC curve (area = {roc_auc:.2f})')
ax_roc.plot([0, 1], [0, 1], color='gray', linestyle='--')
ax_roc.set_xlim([0.0, 1.0])
ax_roc.set_ylim([0.0, 1.05])
ax_roc.set_xlabel('False Positive Rate')
ax_roc.set_ylabel('True Positive Rate')
ax_roc.set_title('Receiver Operating Characteristic (ROC) Curve')
ax_roc.legend(loc="lower right")
plt.close(fig_roc)
return accuracy, roc_score, report, fig, fig_roc
# Batch processing for all models
def process_zip(zip_file):
extracted_dir = extract_zip(zip_file.name)
# Run classification for each model
results = {}
for idx in range(len(models)):
print(f"Processing with model {models[idx]}") # Debugging to show which model is being used
# Create a new pipeline for each model within the loop
pipe = pipeline("image-classification", f"{models[idx]}")
print(f"Initialized pipeline for {models[idx]}") # Confirm pipeline is initialized correctly
# Classify images with the correct pipeline per model
labels, preds, images = classify_images(extracted_dir, pipe, idx)
# Debugging: Print the predictions to ensure they're different
print(f"Predictions for model {models[idx]}: {preds}")
accuracy, roc_score, report, cm_fig, roc_fig = evaluate_model(labels, preds)
# Store results for each model
results[f'Model_{idx}_accuracy'] = accuracy
results[f'Model_{idx}_roc_score'] = roc_score
results[f'Model_{idx}_report'] = report
results[f'Model_{idx}_cm_fig'] = cm_fig
results[f'Model_{idx}_roc_fig'] = roc_fig
shutil.rmtree(extracted_dir) # Clean up extracted files
# Return results for all models
return (results['Model_0_accuracy'], results['Model_0_roc_score'], results['Model_0_report'],
results['Model_0_cm_fig'], results['Model_0_roc_fig'],
results['Model_1_accuracy'], results['Model_1_roc_score'], results['Model_1_report'],
results['Model_1_cm_fig'], results['Model_1_roc_fig'],
results['Model_2_accuracy'], results['Model_2_roc_score'], results['Model_2_report'],
results['Model_2_cm_fig'], results['Model_2_roc_fig'])
# Single image section
def load_url(url):
try:
urllib.request.urlretrieve(f'{url}', f"{uid}tmp_im.png")
image = Image.open(f"{uid}tmp_im.png")
mes = "Image Loaded"
except Exception as e:
image = None
mes = f"Image not Found
Error: {e}"
return image, mes
def tot_prob():
try:
fin_out = sum([result["Real"] for result in fin_sum]) / len(fin_sum)
fin_sub = 1 - fin_out
out = {
"Real": f"{fin_out:.4f}",
"AI": f"{fin_sub:.4f}"
}
return out
except Exception as e:
print(e)
return None
def fin_clear():
fin_sum.clear()
return None
# Set up Gradio app
with gr.Blocks() as app:
gr.Markdown("""AI Image Detector
(Test Demo - accuracy varies by model)
""")
with gr.Tabs():
# Tab for single image detection
with gr.Tab("Single Image Detection"):
with gr.Column():
inp = gr.Image(type='pil')
in_url = gr.Textbox(label="Image URL")
with gr.Row():
load_btn = gr.Button("Load URL")
btn = gr.Button("Detect AI")
mes = gr.HTML("""""")
with gr.Group():
with gr.Row():
fin = gr.Label(label="Final Probability")
with gr.Row():
for i, model in enumerate(models):
with gr.Box():
gr.HTML(f"""Testing on Model {i}: {model}""")
globals()[f'outp{i}'] = gr.HTML("""""")
globals()[f'n_out{i}'] = gr.Label(label="Output")
btn.click(fin_clear, None, fin, show_progress=False)
load_btn.click(load_url, in_url, [inp, mes])
btn.click(aiornot0, [inp], [outp0, n_out0]).then(
aiornot1, [inp], [outp1, n_out1]).then(
aiornot2, [inp], [outp2, n_out2]).then(
tot_prob, None, fin, show_progress=False)
# Tab for batch processing
with gr.Tab("Batch Image Processing"):
zip_file = gr.File(label="Upload Zip (two folders: real, ai)")
batch_btn = gr.Button("Process Batch")
for i, model in enumerate(models):
with gr.Group():
gr.Markdown(f"### Results for {model}")
globals()[f'output_acc{i}'] = gr.Label(label=f"Model {i} Accuracy")
globals()[f'output_roc{i}'] = gr.Label(label=f"Model {i} ROC Score")
globals()[f'output_report{i}'] = gr.Textbox(label=f"Model {i} Classification Report", lines=10)
globals()[f'output_cm{i}'] = gr.Plot(label=f"Model {i} Confusion Matrix")
globals()[f'output_roc_plot{i}'] = gr.Plot(label=f"Model {i} ROC Curve")
# Connect batch processing
batch_btn.click(process_zip, zip_file,
[output_acc0, output_roc0, output_report0, output_cm0, output_roc_plot0,
output_acc1, output_roc1, output_report1, output_cm1, output_roc_plot1,
output_acc2, output_roc2, output_report2, output_cm2, output_roc_plot2])
app.launch(show_api=False, max_threads=24)