V2

app.py

@@ -1,6 +1,6 @@
 import streamlit as st
 from src.preprocessing import prepare_image
-from src.analyzers import run_fft, run_ela, run_resnet, run_siglip
+from src.analyzers import run_fft, run_ela, run_resnet, run_siglip, run_sdxl_detector, run_deepfake_detector
 from src.ensemble import calculate_final_verdict
 
 st.set_page_config(page_title="AI Image Detector", layout="centered", page_icon="👁️")
@@ -34,14 +34,16 @@ st.markdown("<h1 style='text-align: center; color: #f0f2f6; font-weight: 800;'>
 with st.sidebar:
     st.header("Methodology")
     st.write('''
-    This tool uses a 4-method ensemble approach to detect AI-generated images vs. real images:
+    This tool uses a 6-method ensemble approach to detect AI-generated images vs. real images:
     
     1. **FFT (Fast Fourier Transform):** Analyzes the image in the frequency domain to find repeating artifacts common in AI generation (like checkerboard patterns).
     2. **ELA (Error Level Analysis):** Highlights areas of an image that are compressed at different quality levels, which can indicate tampering or non-uniform synthesis.
-    3. **ResNet Model:** Uses `umm-maybe/AI-image-detector` to classify AI vs. Real features using a deep neural network approach.
-    4. **SigLIP Model:** Uses `Ateeqq/ai-vs-human-image-detector` for an additional vision-language-based confidence score.
+    3. **ResNet Model:** Uses `umm-maybe/AI-image-detector`.
+    4. **SigLIP Model:** Uses `Ateeqq/ai-vs-human-image-detector`.
+    5. **SDXL Detector:** Uses `Organika/sdxl-detector` to catch modern diffusion artifacts.
+    6. **DeepFake Detector (ViT):** Uses `prithivMLmods/Deep-Fake-Detector-v2-Model` for state-of-the-art vision transformer detection.
     
-    **The Final Judge:** The final verdict averages the ResNet and SigLIP scores.
+    **The Final Judge:** The final verdict averages all 4 deep-learning AI model scores.
     ''')
     st.markdown("---")
     st.write("Developed for transparency in AI generation.")
@@ -92,11 +94,26 @@ if uploaded_file is not None:
             siglip_score = run_siglip(rgb_img)
         st.metric(label="SigLIP Pipeline ('Fake' Confidence)", value=f"{siglip_score * 100:.2f}%")
         
+    st.markdown("<br>", unsafe_allow_html=True)
+    
+    # Analysis row 3: Additional Advanced Models
+    col5, col6 = st.columns(2)
+    
+    with col5:
+        with st.spinner("Engine 5: Analyzing with SDXL Detector..."):
+            sdxl_score = run_sdxl_detector(rgb_img)
+        st.metric(label="SDXL Detector ('Fake' Confidence)", value=f"{sdxl_score * 100:.2f}%")
+        
+    with col6:
+        with st.spinner("Engine 6: Analyzing with ViT Deepfake Detector..."):
+            deepfake_score = run_deepfake_detector(rgb_img)
+        st.metric(label="ViT Deepfake Detector ('Fake' Confidence)", value=f"{deepfake_score * 100:.2f}%")
+        
     st.markdown("---")
     st.header("The Final Judge")
     
     with st.spinner("Combining scores and formulating final verdict..."):
-        verdict, confidence = calculate_final_verdict(resnet_score, siglip_score)
+        verdict, confidence = calculate_final_verdict(resnet_score, siglip_score, sdxl_score, deepfake_score)
         
     if verdict == "FAKE":
         st.error(f"### Final Verdict: {verdict}")

src/analyzers.py

@@ -13,6 +13,14 @@ def load_resnet_pipeline():
 def load_siglip_pipeline():
     return pipeline("image-classification", model="Ateeqq/ai-vs-human-image-detector")
 
+@st.cache_resource
+def load_sdxl_pipeline():
+    return pipeline("image-classification", model="Organika/sdxl-detector")
+
+@st.cache_resource
+def load_deepfake_pipeline():
+    return pipeline("image-classification", model="prithivMLmods/Deep-Fake-Detector-v2-Model")
+
 def run_fft(grayscale_array):
     """
     Calculates the Fast Fourier Transform magnitude spectrum on the grayscale array.
@@ -88,3 +96,41 @@ def run_siglip(image):
             fake_score = 1.0 - res['score']
             
     return float(fake_score)
+
+def run_sdxl_detector(image):
+    """
+    Uses the Organika/sdxl-detector pipeline.
+    Returns the float confidence score for 'fake'/'AI'.
+    """
+    sdxl_pipeline = load_sdxl_pipeline()
+    results = sdxl_pipeline(image)
+    
+    fake_score = 0.0
+    for res in results:
+        label_lower = res['label'].lower()
+        if label_lower in ['artificial', 'fake', 'ai', 'ai generated']:
+            fake_score = res['score']
+            break
+        elif label_lower in ['human', 'real']:
+            fake_score = 1.0 - res['score']
+            
+    return float(fake_score)
+
+def run_deepfake_detector(image):
+    """
+    Uses the prithivMLmods/Deep-Fake-Detector-v2-Model pipeline.
+    Returns the float confidence score for 'fake'/'Deepfake'.
+    """
+    deepfake_pipeline = load_deepfake_pipeline()
+    results = deepfake_pipeline(image)
+    
+    fake_score = 0.0
+    for res in results:
+        label_lower = res['label'].lower()
+        if label_lower in ['deepfake', 'fake', 'artificial', 'ai']:
+            fake_score = res['score']
+            break
+        elif label_lower in ['realism', 'real', 'human']:
+            fake_score = 1.0 - res['score']
+            
+    return float(fake_score)

src/ensemble.py

@@ -1,9 +1,9 @@
-def calculate_final_verdict(resnet_score, siglip_score):
+def calculate_final_verdict(resnet_score, siglip_score, sdxl_score, deepfake_score):
     """
-    Averages the two AI model scores and returns a final string ("REAL" or "FAKE") 
+    Averages the four AI model scores and returns a final string ("REAL" or "FAKE") 
     alongside the combined percentage confidence.
     """
-    average_score = (resnet_score + siglip_score) / 2.0
+    average_score = (resnet_score + siglip_score + sdxl_score + deepfake_score) / 4.0
     
     # If average "fake" score is > 0.5, it's considered FAKE
     if average_score > 0.5: