major(refactor): prep for refactor project

README.md

@@ -6,17 +6,19 @@ colorFrom: yellow
 colorTo: yellow
 sdk: gradio
 sdk_version: 5.33.0
-app_file: app_optimized.py
+app_file: app.py
 pinned: true
 models:
 - aiwithoutborders-xyz/OpenSight-CommunityForensics-Deepfake-ViT
 license: mit
 tags:
-  - mcp-server-track
+  - deepfake-detection
   - ai-agents
   - leaderboards
-  - incentivized-contests
-  - Agents-MCP-Hackathon
+  - deepfake
+  - detection
+  - ensemble
+  - forensics
 
 ---
 

app_optimized.py → app.backup.7-26.py

@@ -8,12 +8,6 @@ import logging
 import io
 import collections
 import onnxruntime
-import json
-from huggingface_hub import CommitScheduler, hf_hub_download, snapshot_download
-from dotenv import load_dotenv
-import concurrent.futures
-import ast
-import torch
 
 from utils.utils import softmax, augment_image
 from forensics.gradient import gradient_processing
@@ -29,6 +23,10 @@ from utils.registry import register_model, MODEL_REGISTRY, ModelEntry
 from agents.ensemble_weights import ModelWeightManager
 from transformers import pipeline, AutoImageProcessor, SwinForImageClassification, Swinv2ForImageClassification, AutoFeatureExtractor, AutoModelForImageClassification
 from torchvision import transforms
+import torch
+import json
+from huggingface_hub import CommitScheduler
+from dotenv import load_dotenv
 
 logging.basicConfig(level=logging.INFO)
 logger = logging.getLogger(__name__)
@@ -68,25 +66,23 @@ device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
 
 # Model paths and class names (copied from app_mcp.py)
 MODEL_PATHS = {
-    "model_1": "LPX55/detection-model-1-ONNX",
-    "model_2": "LPX55/detection-model-2-ONNX",
-    "model_3": "LPX55/detection-model-3-ONNX",
+    "model_1": "haywoodsloan/ai-image-detector-deploy",
+    "model_2": "Heem2/AI-vs-Real-Image-Detection",
+    "model_3": "Organika/sdxl-detector",
     "model_4": "cmckinle/sdxl-flux-detector_v1.1",
-    "model_5": "LPX55/detection-model-5-ONNX",
-    "model_6": "LPX55/detection-model-6-ONNX",
-    "model_7": "LPX55/detection-model-7-ONNX",
-    "model_8": "aiwithoutborders-xyz/CommunityForensics-DeepfakeDet-ViT"
+    "model_5": "prithivMLmods/Deep-Fake-Detector-v2-Model",
+    "model_6": "ideepankarsharma2003/AI_ImageClassification_MidjourneyV6_SDXL",
+    "model_7": "date3k2/vit-real-fake-classification-v4"
 }
 
 CLASS_NAMES = {
     "model_1": ['artificial', 'real'],
     "model_2": ['AI Image', 'Real Image'],
-    "model_3": ['artificial', 'human'],
+    "model_3": ['AI', 'Real'],
     "model_4": ['AI', 'Real'],
     "model_5": ['Realism', 'Deepfake'],
     "model_6": ['ai_gen', 'human'],
     "model_7": ['Fake', 'Real'],
-    "model_8": ['Fake', 'Real'],
 }
 
 def preprocess_resize_256(image):
@@ -101,7 +97,7 @@ def preprocess_resize_224(image):
 
 def postprocess_pipeline(prediction, class_names):
     # Assumes HuggingFace pipeline output
-    return {pred['label']: float(pred['score']) for pred in prediction}
+    return {pred['label']: pred['score'] for pred in prediction}
 
 def postprocess_logits(outputs, class_names):
     # Assumes model output with logits
@@ -109,362 +105,182 @@ def postprocess_logits(outputs, class_names):
     probabilities = softmax(logits)
     return {class_names[i]: probabilities[i] for i in range(len(class_names))}
 
-def postprocess_binary_output(output, class_names):
-    # output can be a dictionary {"probabilities": numpy_array} or directly a numpy_array
-    probabilities_array = None
-    if isinstance(output, dict) and "probabilities" in output:
-        probabilities_array = output["probabilities"]
-    elif isinstance(output, np.ndarray):
-        probabilities_array = output
-    else:
-        logger.warning(f"Unexpected output type for binary post-processing: {type(output)}. Expected dict with 'probabilities' or numpy.ndarray.")
-        return {class_names[0]: 0.0, class_names[1]: 1.0}
-
-    logger.info(f"Debug: Probabilities array entering postprocess_binary_output: {probabilities_array}, type: {type(probabilities_array)}, shape: {probabilities_array.shape}")
-
-    if probabilities_array is None:
-        logger.warning("Probabilities array is None after extracting from output. Returning default scores.")
-        return {class_names[0]: 0.0, class_names[1]: 1.0}
-
-    if probabilities_array.size == 1:
-        fake_prob = float(probabilities_array.item())
-    elif probabilities_array.size == 2:
-        fake_prob = float(probabilities_array[0])
-    else:
-        logger.warning(f"Unexpected probabilities array shape for binary post-processing: {probabilities_array.shape}. Expected size 1 or 2.")
-        return {class_names[0]: 0.0, class_names[1]: 1.0}
-
-    real_prob = 1.0 - fake_prob # Ensure Fake and Real sum to 1
-    return {class_names[0]: fake_prob, class_names[1]: real_prob}
-
-def infer_gradio_api(image_path):
-    client = Client("aiwithoutborders-xyz/OpenSight-Community-Forensics-Preview")
-    result_dict = client.predict(
-        input_image=handle_file(image_path),
-        api_name="/simple_predict"
-    )
-    logger.info(f"Debug: Raw result_dict from Gradio API (model_8): {result_dict}, type: {type(result_dict)}")
-    # result_dict is already a dictionary, no need for ast.literal_eval
-    fake_probability = result_dict.get('Fake Probability', 0.0)
-    logger.info(f"Debug: Parsed result_dict: {result_dict}, Extracted fake_probability: {fake_probability}")
-    return {"probabilities": np.array([fake_probability])} # Return as a numpy array with one element
-
-# New preprocess function for Gradio API
-def preprocess_gradio_api(image: Image.Image):
-    # The Gradio API expects a file path, so we need to save the PIL Image to a temporary file.
-    temp_file_path = "./temp_gradio_input.png"
-    image.save(temp_file_path)
-    return temp_file_path
-
-# New postprocess function for Gradio API (adapting postprocess_binary_output)
-def postprocess_gradio_api(gradio_output, class_names):
-    # gradio_output is expected to be a dictionary like {"probabilities": np.array([fake_prob])}
-    probabilities_array = None
-    if isinstance(gradio_output, dict) and "probabilities" in gradio_output:
-        probabilities_array = gradio_output["probabilities"]
-    elif isinstance(gradio_output, np.ndarray):
-        probabilities_array = gradio_output
-    else:
-        logger.warning(f"Unexpected output type for Gradio API post-processing: {type(gradio_output)}. Expected dict with 'probabilities' or numpy.ndarray.")
-        return {class_names[0]: 0.0, class_names[1]: 1.0}
-
-    logger.info(f"Debug: Probabilities array entering postprocess_gradio_api: {probabilities_array}, type: {type(probabilities_array)}, shape: {probabilities_array.shape}")
-
-    if probabilities_array is None or probabilities_array.size == 0:
-        logger.warning("Probabilities array is None or empty after extracting from Gradio API output. Returning default scores.")
-        return {class_names[0]: 0.0, class_names[1]: 1.0}
-
-    # It should always be a single element array for fake probability
-    fake_prob = float(probabilities_array.item())
-    real_prob = 1.0 - fake_prob
-
-    return {class_names[0]: fake_prob, class_names[1]: real_prob}
-
 def register_model_with_metadata(model_id, model, preprocess, postprocess, class_names, display_name, contributor, model_path, architecture=None, dataset=None):
     entry = ModelEntry(model, preprocess, postprocess, class_names, display_name=display_name, contributor=contributor, model_path=model_path, architecture=architecture, dataset=dataset)
     MODEL_REGISTRY[model_id] = entry
 
+# Load and register models (copied from app_mcp.py)
+# image_processor_1 = AutoImageProcessor.from_pretrained(MODEL_PATHS["model_1"], use_fast=True)
+# model_1 = Swinv2ForImageClassification.from_pretrained(MODEL_PATHS["model_1"]).to(device)
+# clf_1 = pipeline(model=model_1, task="image-classification", image_processor=image_processor_1, device=device)
+# register_model_with_metadata(
+#     "model_1", clf_1, preprocess_resize_256, postprocess_pipeline, CLASS_NAMES["model_1"],
+#     display_name="SWIN1", contributor="haywoodsloan", model_path=MODEL_PATHS["model_1"],
+#     architecture="SwinV2", dataset="TBA"
+# )
 
-def load_onnx_model_and_preprocessor(hf_model_id):
-    # model_dir = snapshot_download(repo_id=hf_model_id, local_dir_use_symlinks=False)
-    
-    # Create a unique local directory for each ONNX model
-    model_specific_dir = os.path.join("./models", hf_model_id.replace('/', '_'))
-    os.makedirs(model_specific_dir, exist_ok=True)
-
-    # Use hf_hub_download to get specific files into the model-specific directory
-    onnx_model_path = hf_hub_download(repo_id=hf_model_id, filename="model_quantized.onnx", subfolder="onnx", local_dir=model_specific_dir, local_dir_use_symlinks=False)
-
-    # Load preprocessor config
-    preprocessor_config = {}
-    try:
-        preprocessor_config_path = hf_hub_download(repo_id=hf_model_id, filename="preprocessor_config.json", local_dir=model_specific_dir, local_dir_use_symlinks=False)
-        with open(preprocessor_config_path, 'r') as f:
-            preprocessor_config = json.load(f)
-    except Exception as e:
-        logger.warning(f"Could not download or load preprocessor_config.json for {hf_model_id}: {e}")
-    
-    # Load model config for class names if available
-    model_config = {}
-    try:
-        model_config_path = hf_hub_download(repo_id=hf_model_id, filename="config.json", local_dir=model_specific_dir, local_dir_use_symlinks=False)
-        with open(model_config_path, 'r') as f:
-            model_config = json.load(f)
-    except Exception as e:
-        logger.warning(f"Could not download or load config.json for {hf_model_id}: {e}")
-
-    return onnxruntime.InferenceSession(onnx_model_path), preprocessor_config, model_config
-
-
-# Cache for ONNX sessions and preprocessors
-_onnx_model_cache = {}
-
-def get_onnx_model_from_cache(hf_model_id):
-    if hf_model_id not in _onnx_model_cache:
-        logger.info(f"Loading ONNX model and preprocessor for {hf_model_id}...")
-        _onnx_model_cache[hf_model_id] = load_onnx_model_and_preprocessor(hf_model_id)
-    return _onnx_model_cache[hf_model_id]
+# --- ONNX Quantized Model Example ---
+ONNX_QUANTIZED_MODEL_PATH = "./models/model_1_quantized.onnx"
 
-def preprocess_onnx_input(image: Image.Image, preprocessor_config: dict):
-    # Preprocess image for ONNX model based on preprocessor_config
+def preprocess_onnx_input(image: Image.Image):
+    # Preprocess image for ONNX model (e.g., for SwinV2, usually 256x256, normalized)
     if image.mode != 'RGB':
         image = image.convert('RGB')
     
-    # Get image size and normalization values from preprocessor_config or use defaults
-    # Use 'size' for initial resize and 'crop_size' for center cropping
-    initial_resize_size = preprocessor_config.get('size', {'height': 224, 'width': 224})
-    crop_size = preprocessor_config.get('crop_size', initial_resize_size['height'])
-    mean = preprocessor_config.get('image_mean', [0.485, 0.456, 0.406])
-    std = preprocessor_config.get('image_std', [0.229, 0.224, 0.225])
-
     transform = transforms.Compose([
-        transforms.Resize((initial_resize_size['height'], initial_resize_size['width'])),
-        transforms.CenterCrop(crop_size), # Apply center crop
+        transforms.Resize((256, 256)),
         transforms.ToTensor(),
-        transforms.Normalize(mean=mean, std=std),
+        transforms.Normalize(mean=[0.485,0.456,0.406], std=[0.229,0.224,0.225]), # ImageNet normalization
     ])
     input_tensor = transform(image)
     # ONNX expects numpy array with batch dimension (1, C, H, W)
     return input_tensor.unsqueeze(0).cpu().numpy()
 
-def infer_onnx_model(hf_model_id, preprocessed_image_np, model_config: dict):
+def infer_onnx_model(preprocessed_image_np):
     try:
-        ort_session, _, _ = get_onnx_model_from_cache(hf_model_id)
-        
-        # Debug: Print expected input shape from ONNX model
-        for input_meta in ort_session.get_inputs():
-            logger.info(f"Debug: ONNX model expected input name: {input_meta.name}, shape: {input_meta.shape}, type: {input_meta.type}")
+        # Ensure the ONNX model exists before trying to load it
+        if not os.path.exists(ONNX_QUANTIZED_MODEL_PATH):
+            logger.error(f"ONNX quantized model not found at: {ONNX_QUANTIZED_MODEL_PATH}")
+            raise FileNotFoundError(f"ONNX quantized model not found at: {ONNX_QUANTIZED_MODEL_PATH}")
 
-        logger.info(f"Debug: preprocessed_image_np shape: {preprocessed_image_np.shape}")
+        ort_session = onnxruntime.InferenceSession(ONNX_QUANTIZED_MODEL_PATH)
         ort_inputs = {ort_session.get_inputs()[0].name: preprocessed_image_np}
         ort_outputs = ort_session.run(None, ort_inputs)
         
+        # Assuming the output is logits, apply softmax to get probabilities
         logits = ort_outputs[0]
-        logger.info(f"Debug: logits type: {type(logits)}, shape: {logits.shape}")
-        # If the model outputs a single logit (e.g., shape (1,)), use .item() to convert to scalar
-        # Otherwise, assume it's a batch of logits (e.g., shape (1, num_classes)) and take the first element (batch dim)
-        # The num_classes in config.json can be misleading; rely on actual output shape.
-        
-        # Apply softmax to the logits to get probabilities for the classes
-        # The softmax function in utils/utils.py now ensures a list of floats
-        probabilities = softmax(logits[0]) # Assuming logits[0] is the relevant output for a single prediction
-
+        probabilities = softmax(logits[0]) # Remove batch dim, apply softmax
         return {"logits": logits, "probabilities": probabilities}
 
     except Exception as e:
-        logger.error(f"Error during ONNX inference for {hf_model_id}: {e}")
+        logger.error(f"Error during ONNX inference: {e}")
         # Return a structure consistent with other model errors
         return {"logits": np.array([]), "probabilities": np.array([])}
 
-def postprocess_onnx_output(onnx_output, model_config):
-    # Get class names from model_config
-    # Prioritize id2label, then check num_classes, otherwise default
-    class_names_map = model_config.get('id2label')
-    if class_names_map:
-        class_names = [class_names_map[k] for k in sorted(class_names_map.keys())]
-    elif model_config.get('num_classes') == 1: # Handle models that output a single value (e.g., probability of 'Fake')
-        class_names = ['Fake', 'Real'] # Assume first class is 'Fake' and second 'Real'
-    else:
-        class_names = {0: 'Fake', 1: 'Real'} # Default to Fake/Real if not found or not 1 class
-        class_names = [class_names[i] for i in sorted(class_names.keys())]
-
+def postprocess_onnx_output(onnx_output, class_names):
     probabilities = onnx_output.get("probabilities")
-    
-    if probabilities is not None:
-        if model_config.get('num_classes') == 1 and len(probabilities) == 2: # Special handling for single output models
-            # The single output is the probability of the 'Fake' class
-            fake_prob = float(probabilities[0])
-            real_prob = float(probabilities[1])
-            return {class_names[0]: fake_prob, class_names[1]: real_prob}
-        elif len(probabilities) == len(class_names):
-            return {class_names[i]: float(probabilities[i]) for i in range(len(class_names))}
-        else:
-            logger.warning("ONNX post-processing: Probabilities length mismatch with class names.")
-            return {name: 0.0 for name in class_names}
+    if probabilities is not None and len(probabilities) == len(class_names):
+        return {class_names[i]: probabilities[i] for i in range(len(class_names))}
     else:
-        logger.warning("ONNX post-processing failed: 'probabilities' key not found in output.")
+        logger.warning("ONNX post-processing failed or class names mismatch.")
         return {name: 0.0 for name in class_names}
 
 # Register the ONNX quantized model
-# Dummy entry for ONNX model to be loaded dynamically
-# We will now register a 'wrapper' that handles dynamic loading
-
-class ONNXModelWrapper:
-    def __init__(self, hf_model_id):
-        self.hf_model_id = hf_model_id
-        self._session = None
-        self._preprocessor_config = None
-        self._model_config = None
-
-    def load(self):
-        if self._session is None:
-            self._session, self._preprocessor_config, self._model_config = get_onnx_model_from_cache(self.hf_model_id)
-            logger.info(f"ONNX model {self.hf_model_id} loaded into wrapper.")
-
-    def __call__(self, image_np):
-        self.load() # Ensure model is loaded on first call
-        # Pass model_config to infer_onnx_model
-        return infer_onnx_model(self.hf_model_id, image_np, self._model_config)
-
-    def preprocess(self, image: Image.Image):
-        self.load()
-        return preprocess_onnx_input(image, self._preprocessor_config)
-
-    def postprocess(self, onnx_output: dict, class_names_from_registry: list): # class_names_from_registry is ignored
-        self.load()
-        return postprocess_onnx_output(onnx_output, self._model_config)
-
-# Consolidate all model loading and registration
-for model_key, hf_model_path in MODEL_PATHS.items():
-    logger.debug(f"Attempting to register model: {model_key} with path: {hf_model_path}")
-    model_num = model_key.replace("model_", "").upper()
-    contributor = "Unknown"
-    architecture = "Unknown"
-    dataset = "TBA"
-
-    current_class_names = CLASS_NAMES.get(model_key, [])
-
-    # Logic for ONNX models (1, 2, 3, 5, 6, 7)
-    if "ONNX" in hf_model_path:
-        logger.debug(f"Model {model_key} identified as ONNX.")
-        logger.info(f"Registering ONNX model: {model_key} from {hf_model_path}")
-        onnx_wrapper_instance = ONNXModelWrapper(hf_model_path)
-
-        # Attempt to derive contributor, architecture, dataset based on model_key
-        if model_key == "model_1":
-            contributor = "haywoodsloan"
-            architecture = "SwinV2"
-            dataset = "DeepFakeDetection"
-        elif model_key == "model_2":
-            contributor = "Heem2"
-            architecture = "ViT"
-            dataset = "DeepFakeDetection"
-        elif model_key == "model_3":
-            contributor = "Organika"
-            architecture = "VIT"
-            dataset = "SDXL"
-        elif model_key == "model_5":
-            contributor = "prithivMLmods"
-            architecture = "VIT"
-        elif model_key == "model_6":
-            contributor = "ideepankarsharma2003"
-            architecture = "SWINv1"
-            dataset = "SDXL, Midjourney"
-        elif model_key == "model_7":
-            contributor = "date3k2"
-            architecture = "VIT"
-
-        display_name_parts = [model_num]
-        if architecture and architecture not in ["Unknown"]:
-            display_name_parts.append(architecture)
-        if dataset and dataset not in ["TBA"]:
-            display_name_parts.append(dataset)
-        display_name = "-".join(display_name_parts)
-        display_name += "_ONNX" # Always append _ONNX for ONNX models
-
-        register_model_with_metadata(
-            model_id=model_key,
-            model=onnx_wrapper_instance, # The callable wrapper for the ONNX model
-            preprocess=onnx_wrapper_instance.preprocess,
-            postprocess=onnx_wrapper_instance.postprocess,
-            class_names=current_class_names, # Initial class names; will be overridden by model_config if available
-            display_name=display_name,
-            contributor=contributor,
-            model_path=hf_model_path,
-            architecture=architecture,
-            dataset=dataset
-        )
-    # Logic for Gradio API model (model_8)
-    elif model_key == "model_8":
-        logger.debug(f"Model {model_key} identified as Gradio API.")
-        logger.info(f"Registering Gradio API model: {model_key} from {hf_model_path}")
-        contributor = "aiwithoutborders-xyz"
-        architecture = "ViT"
-        dataset = "DeepfakeDetection"
-
-        display_name_parts = [model_num]
-        if architecture and architecture not in ["Unknown"]:
-            display_name_parts.append(architecture)
-        if dataset and dataset not in ["TBA"]:
-            display_name_parts.append(dataset)
-        display_name = "-".join(display_name_parts)
-
-        register_model_with_metadata(
-            model_id=model_key,
-            model=infer_gradio_api,
-            preprocess=preprocess_gradio_api,
-            postprocess=postprocess_gradio_api,
-            class_names=current_class_names,
-            display_name=display_name,
-            contributor=contributor,
-            model_path=hf_model_path,
-            architecture=architecture,
-            dataset=dataset
-        )
-    # Logic for PyTorch/Hugging Face pipeline models (currently only model_4)
-    elif model_key == "model_4": # Explicitly handle model_4
-        logger.debug(f"Model {model_key} identified as PyTorch/HuggingFace pipeline.")
-        logger.info(f"Registering HuggingFace pipeline/AutoModel: {model_key} from {hf_model_path}")
-        contributor = "cmckinle"
-        architecture = "VIT"
-        dataset = "SDXL, FLUX"
-
-        display_name_parts = [model_num]
-        if architecture and architecture not in ["Unknown"]:
-            display_name_parts.append(architecture)
-        if dataset and dataset not in ["TBA"]:
-            display_name_parts.append(dataset)
-        display_name = "-".join(display_name_parts)
-
-        current_processor = AutoFeatureExtractor.from_pretrained(hf_model_path, device=device)
-        model_instance = AutoModelForImageClassification.from_pretrained(hf_model_path).to(device)
-
-        preprocess_func = preprocess_resize_256
-        postprocess_func = postprocess_logits
-
-        def custom_infer(image, processor_local=current_processor, model_local=model_instance):
-            inputs = processor_local(image, return_tensors="pt").to(device)
-            with torch.no_grad():
-                outputs = model_local(**inputs)
-            return outputs
-        model_instance = custom_infer
-
-        register_model_with_metadata(
-            model_id=model_key,
-            model=model_instance,
-            preprocess=preprocess_func,
-            postprocess=postprocess_func,
-            class_names=current_class_names,
-            display_name=display_name,
-            contributor=contributor,
-            model_path=hf_model_path,
-            architecture=architecture,
-            dataset=dataset
-        )
-    else: # Fallback for any unhandled models (shouldn't happen if MODEL_PATHS is fully covered)
-        logger.warning(f"Could not automatically load and register model: {model_key} from {hf_model_path}. No matching registration logic found.")
+register_model_with_metadata(
+    "model_1_onnx_quantized", 
+    infer_onnx_model, 
+    preprocess_onnx_input, 
+    postprocess_onnx_output, 
+    CLASS_NAMES["model_1"], # Assuming it uses the same class names as model_1
+    display_name="SWIN1", 
+    contributor="haywoodsloan", 
+    model_path=ONNX_QUANTIZED_MODEL_PATH,
+    architecture="SwinV2", 
+    dataset="TBA"
+)
+# --- End ONNX Quantized Model Example ---
 
+clf_2 = pipeline("image-classification", model=MODEL_PATHS["model_2"], device=device)
+register_model_with_metadata(
+    "model_2", clf_2, preprocess_resize_224, postprocess_pipeline, CLASS_NAMES["model_2"],
+    display_name="VIT2", contributor="Heem2", model_path=MODEL_PATHS["model_2"],
+    architecture="ViT", dataset="TBA"
+)
 
+feature_extractor_3 = AutoFeatureExtractor.from_pretrained(MODEL_PATHS["model_3"], device=device)
+model_3 = AutoModelForImageClassification.from_pretrained(MODEL_PATHS["model_3"]).to(device)
+def preprocess_256(image):
+    if image.mode != 'RGB':
+        image = image.convert('RGB')
+    return transforms.Resize((256, 256))(image)
+def postprocess_logits_model3(outputs, class_names):
+    logits = outputs.logits.cpu().numpy()[0]
+    probabilities = softmax(logits)
+    return {class_names[i]: probabilities[i] for i in range(len(class_names))}
+def model3_infer(image):
+    inputs = feature_extractor_3(image, return_tensors="pt").to(device)
+    with torch.no_grad():
+        outputs = model_3(**inputs)
+    return outputs
+register_model_with_metadata(
+    "model_3", model3_infer, preprocess_256, postprocess_logits_model3, CLASS_NAMES["model_3"],
+    display_name="SDXL3", contributor="Organika", model_path=MODEL_PATHS["model_3"],
+    architecture="VIT", dataset="SDXL"
+)
+
+feature_extractor_4 = AutoFeatureExtractor.from_pretrained(MODEL_PATHS["model_4"], device=device)
+model_4 = AutoModelForImageClassification.from_pretrained(MODEL_PATHS["model_4"]).to(device)
+def model4_infer(image):
+    inputs = feature_extractor_4(image, return_tensors="pt").to(device)
+    with torch.no_grad():
+        outputs = model_4(**inputs)
+    return outputs
+def postprocess_logits_model4(outputs, class_names):
+    logits = outputs.logits.cpu().numpy()[0]
+    probabilities = softmax(logits)
+    return {class_names[i]: probabilities[i] for i in range(len(class_names))}
+register_model_with_metadata(
+    "model_4", model4_infer, preprocess_256, postprocess_logits_model4, CLASS_NAMES["model_4"],
+    display_name="XLFLUX4", contributor="cmckinle", model_path=MODEL_PATHS["model_4"],
+    architecture="VIT", dataset="SDXL, FLUX"
+)
+
+clf_5 = pipeline("image-classification", model=MODEL_PATHS["model_5"], device=device)
+register_model_with_metadata(
+    "model_5", clf_5, preprocess_resize_224, postprocess_pipeline, CLASS_NAMES["model_5"],
+    display_name="VIT5", contributor="prithivMLmods", model_path=MODEL_PATHS["model_5"],
+    architecture="VIT", dataset="TBA"
+)
+
+image_processor_6 = AutoImageProcessor.from_pretrained(MODEL_PATHS["model_6"], use_fast=True)
+model_6 = SwinForImageClassification.from_pretrained(MODEL_PATHS["model_6"]).to(device)
+clf_6 = pipeline(model=model_6, task="image-classification", image_processor=image_processor_6, device=device)
+register_model_with_metadata(
+    "model_6", clf_6, preprocess_resize_224, postprocess_pipeline, CLASS_NAMES["model_6"],
+    display_name="SWIN6", contributor="ideepankarsharma2003", model_path=MODEL_PATHS["model_6"],
+    architecture="SWINv1", dataset="SDXL, Midjourney"
+)
+
+image_processor_7 = AutoImageProcessor.from_pretrained(MODEL_PATHS["model_7"], use_fast=True)
+model_7 = AutoModelForImageClassification.from_pretrained(MODEL_PATHS["model_7"]).to(device)
+clf_7 = pipeline(model=model_7, task="image-classification", image_processor=image_processor_7, device=device)
+register_model_with_metadata(
+    "model_7", clf_7, preprocess_resize_224, postprocess_pipeline, CLASS_NAMES["model_7"],
+    display_name="VIT7", contributor="date3k2", model_path=MODEL_PATHS["model_7"],
+    architecture="VIT", dataset="TBA"
+)
+
+# def postprocess_simple_prediction(result, class_names):
+#     scores = {name: 0.0 for name in class_names}
+#     fake_prob = result.get("Fake Probability")
+#     if fake_prob is not None:
+#         # Assume class_names = ["AI", "REAL"]
+#         scores["AI"] = float(fake_prob)
+#         scores["REAL"] = 1.0 - float(fake_prob)
+#     return scores
+
+# def simple_prediction(img):
+#     client = Client("aiwithoutborders-xyz/OpenSight-Community-Forensics-Preview")
+#     client.view_api()
+#     print(type(img))
+#     result = client.predict(
+#            handle_file(img),
+#            api_name="simple_predict"
+#     )
+#     return result
+
+
+# register_model_with_metadata(
+#     model_id="simple_prediction",
+#     model=simple_prediction,
+#     preprocess=None,
+#     postprocess=postprocess_simple_prediction,
+#     class_names=["AI", "REAL"], 
+#     display_name="Community Forensics",
+#     contributor="Jeongsoo Park",
+#     model_path="aiwithoutborders-xyz/CommunityForensics-DeepfakeDet-ViT",
+#     architecture="ViT", dataset="GOAT"
+# )
 def infer(image: Image.Image, model_id: str, confidence_threshold: float = 0.75) -> dict:
     """Predict using a specific model.
 
@@ -481,14 +297,8 @@ def infer(image: Image.Image, model_id: str, confidence_threshold: float = 0.75)
     try:
         result = entry.model(img)
         scores = entry.postprocess(result, entry.class_names)
-        
-        def _to_float_scalar(value):
-            if isinstance(value, np.ndarray):
-                return float(value.item()) # Convert numpy array scalar to Python float
-            return float(value) # Already a Python scalar or convertible type
-
-        ai_score = _to_float_scalar(scores.get(entry.class_names[0], 0.0))
-        real_score = _to_float_scalar(scores.get(entry.class_names[1], 0.0))
+        ai_score = float(scores.get(entry.class_names[0], 0.0))
+        real_score = float(scores.get(entry.class_names[1], 0.0))
         label = "AI" if ai_score >= confidence_threshold else ("REAL" if real_score >= confidence_threshold else "UNCERTAIN")
         return {
             "Model": entry.display_name,
@@ -561,97 +371,29 @@ def full_prediction(img, confidence_threshold, rotate_degrees, noise_level, shar
     results = []
     table_rows = []
 
-    # Initialize lists for forensic outputs, starting with the original augmented image
-    cleaned_forensics_images = []
-    forensic_output_descriptions = []
-
-    # Always add the original augmented image first for forensic display
-    if isinstance(img_pil, Image.Image):
-        cleaned_forensics_images.append(img_pil)
-        forensic_output_descriptions.append(f"Original augmented image (PIL): {img_pil.width}x{img_pil.height}")
-    elif isinstance(img_pil, np.ndarray):
-        try:
-            pil_img_from_np = Image.fromarray(img_pil)
-            cleaned_forensics_images.append(pil_img_from_np)
-            forensic_output_descriptions.append(f"Original augmented image (numpy converted to PIL): {pil_img_from_np.width}x{pil_img_from_np.height}")
-        except Exception as e:
-            logger.warning(f"Could not convert original numpy image to PIL for gallery: {e}")
-    
-    # Yield initial state with augmented image and empty model predictions
-    yield img_pil, cleaned_forensics_images, table_rows, "[]", "<div style='font-size: 2.2em; font-weight: bold;padding: 10px;'>Consensus: <span style='color:orange'>UNCERTAIN</span></div>"
-
-
     # Stream results as each model finishes
     for model_id in MODEL_REGISTRY:
         model_start = time.time()
         result = infer(img_pil, model_id, confidence_threshold)
         model_end = time.time()
-
-        # Helper to ensure values are Python floats, handling numpy scalars
-        def _ensure_float_scalar(value):
-            if isinstance(value, np.ndarray):
-                return float(value.item()) # Convert numpy array scalar to Python float
-            return float(value) # Already a Python scalar or convertible type
-
-        ai_score_val = _ensure_float_scalar(result.get("AI Score", 0.0))
-        real_score_val = _ensure_float_val = _ensure_float_scalar(result.get("Real Score", 0.0))
-
         monitor_agent.monitor_prediction(
             model_id,
             result["Label"],
-            max(ai_score_val, real_score_val),
+            max(result.get("AI Score", 0.0), result.get("Real Score", 0.0)),
             model_end - model_start
         )
         model_predictions_raw[model_id] = result
-        confidence_scores[model_id] = max(ai_score_val, real_score_val)
+        confidence_scores[model_id] = max(result.get("AI Score", 0.0), result.get("Real Score", 0.0))
         results.append(result)
         table_rows.append([
             result.get("Model", ""),
             result.get("Contributor", ""),
-            round(ai_score_val, 5),
-            round(real_score_val, 5),
+            round(result.get("AI Score", 0.0), 3) if result.get("AI Score") is not None else 0.0,
+            round(result.get("Real Score", 0.0), 3) if result.get("Real Score") is not None else 0.0,
             result.get("Label", "Error")
         ])
         # Yield partial results: only update the table, others are None
-        yield None, cleaned_forensics_images, table_rows, None, None # Keep cleaned_forensics_images as is (only augmented image for now)
-
-    # Multi-threaded forensic processing
-    def _run_forensic_task(task_func, img_input, description, **kwargs):
-        try:
-            result_img = task_func(img_input, **kwargs)
-            return result_img, description
-        except Exception as e:
-            logger.error(f"Error processing forensic task {task_func.__name__}: {e}")
-            return None, f"Error processing {description}: {str(e)}"
-
-    with concurrent.futures.ThreadPoolExecutor() as executor:
-        future_ela1 = executor.submit(_run_forensic_task, ELA, img_np_og, "ELA analysis (Pass 1): Grayscale error map, quality 75.", quality=75, scale=50, contrast=20, linear=False, grayscale=True)
-        future_ela2 = executor.submit(_run_forensic_task, ELA, img_np_og, "ELA analysis (Pass 2): Grayscale error map, quality 75, enhanced contrast.", quality=75, scale=75, contrast=25, linear=False, grayscale=True)
-        future_ela3 = executor.submit(_run_forensic_task, ELA, img_np_og, "ELA analysis (Pass 3): Color error map, quality 75, enhanced contrast.", quality=75, scale=75, contrast=25, linear=False, grayscale=False)
-        future_gradient1 = executor.submit(_run_forensic_task, gradient_processing, img_np_og, "Gradient processing: Highlights edges and transitions.")
-        future_gradient2 = executor.submit(_run_forensic_task, gradient_processing, img_np_og, "Gradient processing: Int=45, Equalize=True", intensity=45, equalize=True)
-        future_minmax1 = executor.submit(_run_forensic_task, minmax_process, img_np_og, "MinMax processing: Deviations in local pixel values.")
-        future_minmax2 = executor.submit(_run_forensic_task, minmax_process, img_np_og, "MinMax processing (Radius=6): Deviations in local pixel values.", radius=6)
-
-        forensic_futures = [future_ela1, future_ela2, future_ela3, future_gradient1, future_gradient2, future_minmax1, future_minmax2]
-        
-        for future in concurrent.futures.as_completed(forensic_futures):
-            processed_img, description = future.result()
-            if processed_img is not None:
-                if isinstance(processed_img, Image.Image):
-                    cleaned_forensics_images.append(processed_img)
-                elif isinstance(processed_img, np.ndarray):
-                    try:
-                        cleaned_forensics_images.append(Image.fromarray(processed_img))
-                    except Exception as e:
-                        logger.warning(f"Could not convert numpy array to PIL Image for gallery: {e}")
-                else:
-                    logger.warning(f"Unexpected type in processed_img from {description}: {type(processed_img)}. Skipping.")
-                
-                forensic_output_descriptions.append(description) # Keep track of descriptions for anomaly agent
-                
-                # Yield partial results: update gallery
-                yield None, cleaned_forensics_images, table_rows, None, None
+        yield None, None, table_rows, None, None
 
     # After all models, compute the rest as before
     image_data_for_context = {
@@ -659,17 +401,6 @@ def full_prediction(img, confidence_threshold, rotate_degrees, noise_level, shar
         "height": img.height,
         "mode": img.mode,
     }
-    forensic_output_descriptions = [
-        f"Original augmented image (PIL): {img_pil.width}x{img_pil.height}",
-        "ELA analysis (Pass 1): Grayscale error map, quality 75.",
-        "ELA analysis (Pass 2): Grayscale error map, quality 75, enhanced contrast.",
-        "ELA analysis (Pass 3): Color error map, quality 75, enhanced contrast.",
-        "Gradient processing: Highlights edges and transitions.",
-        "Gradient processing: Int=45, Equalize=True",
-        "MinMax processing: Deviations in local pixel values.",
-        "MinMax processing (Radius=6): Deviations in local pixel values.",
-        # "Bit Plane extractor: Visualization of individual bit planes from different color channels."
-    ]
     detected_context_tags = context_agent.infer_context_tags(image_data_for_context, model_predictions_raw)
     logger.info(f"Detected context tags: {detected_context_tags}")
     adjusted_weights = weight_manager.adjust_weights(model_predictions_raw, confidence_scores, context_tags=detected_context_tags)
@@ -686,26 +417,26 @@ def full_prediction(img, confidence_threshold, rotate_degrees, noise_level, shar
     elif weighted_predictions["REAL"] > weighted_predictions["AI"] and weighted_predictions["REAL"] > weighted_predictions["UNCERTAIN"]:
         final_prediction_label = "REAL"
     optimization_agent.analyze_performance(final_prediction_label, None)
-    # gradient_image = gradient_processing(img_np_og)
-    # gradient_image2 = gradient_processing(img_np_og, intensity=45, equalize=True)
-    # minmax_image = minmax_process(img_np_og)
-    # minmax_image2 = minmax_process(img_np_og, radius=6)
-    # # bitplane_image = bit_plane_extractor(img_pil)
-    # ela1 = ELA(img_np_og, quality=75, scale=50, contrast=20, linear=False, grayscale=True)
-    # ela2 = ELA(img_np_og, quality=75, scale=75, contrast=25, linear=False, grayscale=True)
-    # ela3 = ELA(img_np_og, quality=75, scale=75, contrast=25, linear=False, grayscale=False)
-    # forensics_images = [img_pil, ela1, ela2, ela3, gradient_image, gradient_image2, minmax_image, minmax_image2]
-    # forensic_output_descriptions = [
-    #     f"Original augmented image (PIL): {img_pil.width}x{img_pil.height}",
-    #     "ELA analysis (Pass 1): Grayscale error map, quality 75.",
-    #     "ELA analysis (Pass 2): Grayscale error map, quality 75, enhanced contrast.",
-    #     "ELA analysis (Pass 3): Color error map, quality 75, enhanced contrast.",
-    #     "Gradient processing: Highlights edges and transitions.",
-    #     "Gradient processing: Int=45, Equalize=True",
-    #     "MinMax processing: Deviations in local pixel values.",
-    #     "MinMax processing (Radius=6): Deviations in local pixel values.",
-    #     # "Bit Plane extractor: Visualization of individual bit planes from different color channels."
-    # ]
+    gradient_image = gradient_processing(img_np_og)
+    gradient_image2 = gradient_processing(img_np_og, intensity=45, equalize=True)
+    minmax_image = minmax_process(img_np_og)
+    minmax_image2 = minmax_process(img_np_og, radius=6)
+    # bitplane_image = bit_plane_extractor(img_pil)
+    ela1 = ELA(img_np_og, quality=75, scale=50, contrast=20, linear=False, grayscale=True)
+    ela2 = ELA(img_np_og, quality=75, scale=75, contrast=25, linear=False, grayscale=True)
+    ela3 = ELA(img_np_og, quality=75, scale=75, contrast=25, linear=False, grayscale=False)
+    forensics_images = [img_pil, ela1, ela2, ela3, gradient_image, gradient_image2, minmax_image, minmax_image2]
+    forensic_output_descriptions = [
+        f"Original augmented image (PIL): {img_pil.width}x{img_pil.height}",
+        "ELA analysis (Pass 1): Grayscale error map, quality 75.",
+        "ELA analysis (Pass 2): Grayscale error map, quality 75, enhanced contrast.",
+        "ELA analysis (Pass 3): Color error map, quality 75, enhanced contrast.",
+        "Gradient processing: Highlights edges and transitions.",
+        "Gradient processing: Int=45, Equalize=True",
+        "MinMax processing: Deviations in local pixel values.",
+        "MinMax processing (Radius=6): Deviations in local pixel values.",
+        # "Bit Plane extractor: Visualization of individual bit planes from different color channels."
+    ]
     anomaly_detection_results = anomaly_agent.analyze_forensic_outputs(forensic_output_descriptions)
     logger.info(f"Forensic anomaly detection: {anomaly_detection_results['summary']}")
     consensus_html = f"<div style='font-size: 2.2em; font-weight: bold;padding: 10px;'>Consensus: <span style='color:{'red' if final_prediction_label == 'AI' else ('green' if final_prediction_label == 'REAL' else 'orange')}'>{final_prediction_label}</span></div>"
@@ -743,11 +474,21 @@ def full_prediction(img, confidence_threshold, rotate_degrees, noise_level, shar
         inference_params=inference_params,
         model_predictions=results,
         ensemble_output=ensemble_output_data,
-        forensic_images=cleaned_forensics_images, # Use the incrementally built list
+        forensic_images=forensics_images,
         agent_monitoring_data=agent_monitoring_data_log,
         human_feedback=None
     )
-
+    cleaned_forensics_images = []
+    for f_img in forensics_images:
+        if isinstance(f_img, Image.Image):
+            cleaned_forensics_images.append(f_img)
+        elif isinstance(f_img, np.ndarray):
+            try:
+                cleaned_forensics_images.append(Image.fromarray(f_img))
+            except Exception as e:
+                logger.warning(f"Could not convert numpy array to PIL Image for gallery: {e}")
+        else:
+            logger.warning(f"Unexpected type in forensic_images: {type(f_img)}. Skipping.")
     logger.info(f"Cleaned forensic images types: {[type(img) for img in cleaned_forensics_images]}")
     for i, res_dict in enumerate(results):
         for key in ["AI Score", "Real Score"]:
@@ -968,7 +709,7 @@ demo = gr.TabbedInterface(
 
 )
 footerMD = """
-## ⚠️ ENSEMBLE TEAM IN TRAINING ⚠️ \n\n
+### ⚠️ ENSEMBLE TEAM IN TRAINING ⚠️ \n\n
 
 1. **DISCLAIMER: METADATA AS WELL AS MEDIA SUBMITTED TO THIS SPACE MAY BE VIEWED AND SELECTED FOR FUTURE DATASETS, PLEASE DO NOT SUBMIT PERSONAL CONTENT. FOR UNTRACKED, PRIVATE USE OF THE MODELS YOU MAY STILL USE [THE ORIGINAL SPACE HERE](https://huggingface.co/spaces/aiwithoutborders-xyz/OpenSight-Deepfake-Detection-Models-Playground), SOTA MODEL INCLUDED.**
 2. **UPDATE 6-13-25**: APOLOGIES FOR THE CONFUSION, WE ARE WORKING TO REVERT THE ORIGINAL REPO BACK TO ITS NON-DATA COLLECTION STATE -- ONLY THE "SIMPLE PREDICTION" ENDPOINT IS CURRENTLY 100% PRIVATE. PLEASE STAY TUNED AS WE FIGURE OUT A SOLUTION FOR THE ENSEMBLE + AGENT TEAM ENDPOINT. IT CAN GET RESOURCE INTENSIVE TO RUN A FULL PREDICTION. ALTERNATIVELY, WE **ENCOURAGE** ANYONE TO FORK AND CONTRIBUTE TO THE PROJECT.
@@ -977,7 +718,7 @@ footerMD = """
 TO SUMMARIZE: DATASET COLLECTION WILL CONTINUE FOR OUR NOVEL ENSEMBLE-TEAM PREDICTION PIPELINE UNTIL WE CAN GET THINGS SORTED OUT. FOR THOSE THAT WISH TO OPT-OUT, WE OFFER THE SIMPLE, BUT [MOST POWERFUL DETECTION MODEL HERE.](https://huggingface.co/spaces/aiwithoutborders-xyz/OpenSight-Community-Forensics-Preview)
 
 """
-footer = gr.Markdown(footerMD, elem_classes="footer")
+footer = gr.Markdown("", elem_classes="footer")
 
 with gr.Blocks() as app:
     demo.render()

app.py

@@ -8,8 +8,14 @@ import logging
 import io
 import collections
 import onnxruntime
+import json
+from huggingface_hub import CommitScheduler, hf_hub_download, snapshot_download
+from dotenv import load_dotenv
+import concurrent.futures
+import ast
+import torch
 
-from utils.utils import softmax, augment_image
+from utils.utils import softmax, augment_image, preprocess_resize_256, preprocess_resize_224, postprocess_pipeline, postprocess_logits, postprocess_binary_output, to_float_scalar
 from forensics.gradient import gradient_processing
 from forensics.minmax import minmax_process
 from forensics.ela import ELA
@@ -23,10 +29,6 @@ from utils.registry import register_model, MODEL_REGISTRY, ModelEntry
 from agents.ensemble_weights import ModelWeightManager
 from transformers import pipeline, AutoImageProcessor, SwinForImageClassification, Swinv2ForImageClassification, AutoFeatureExtractor, AutoModelForImageClassification
 from torchvision import transforms
-import torch
-import json
-from huggingface_hub import CommitScheduler
-from dotenv import load_dotenv
 
 logging.basicConfig(level=logging.INFO)
 logger = logging.getLogger(__name__)
@@ -66,221 +68,356 @@ device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
 
 # Model paths and class names (copied from app_mcp.py)
 MODEL_PATHS = {
-    "model_1": "haywoodsloan/ai-image-detector-deploy",
-    "model_2": "Heem2/AI-vs-Real-Image-Detection",
-    "model_3": "Organika/sdxl-detector",
+    "model_1": "LPX55/detection-model-1-ONNX",
+    "model_2": "LPX55/detection-model-2-ONNX",
+    "model_3": "LPX55/detection-model-3-ONNX",
     "model_4": "cmckinle/sdxl-flux-detector_v1.1",
-    "model_5": "prithivMLmods/Deep-Fake-Detector-v2-Model",
-    "model_6": "ideepankarsharma2003/AI_ImageClassification_MidjourneyV6_SDXL",
-    "model_7": "date3k2/vit-real-fake-classification-v4"
+    "model_5": "LPX55/detection-model-5-ONNX",
+    "model_6": "LPX55/detection-model-6-ONNX",
+    "model_7": "LPX55/detection-model-7-ONNX",
+    "model_8": "aiwithoutborders-xyz/CommunityForensics-DeepfakeDet-ViT"
 }
 
 CLASS_NAMES = {
     "model_1": ['artificial', 'real'],
     "model_2": ['AI Image', 'Real Image'],
-    "model_3": ['AI', 'Real'],
+    "model_3": ['artificial', 'human'],
     "model_4": ['AI', 'Real'],
     "model_5": ['Realism', 'Deepfake'],
     "model_6": ['ai_gen', 'human'],
     "model_7": ['Fake', 'Real'],
+    "model_8": ['Fake', 'Real'],
 }
 
-def preprocess_resize_256(image):
-    if image.mode != 'RGB':
-        image = image.convert('RGB')
-    return transforms.Resize((256, 256))(image)
 
-def preprocess_resize_224(image):
-    if image.mode != 'RGB':
-        image = image.convert('RGB')
-    return transforms.Resize((224, 224))(image)
+def infer_gradio_api(image_path):
+    client = Client("aiwithoutborders-xyz/OpenSight-Community-Forensics-Preview")
+    result_dict = client.predict(
+        input_image=handle_file(image_path),
+        api_name="/simple_predict"
+    )
+    logger.info(f"Debug: Raw result_dict from Gradio API (model_8): {result_dict}, type: {type(result_dict)}")
+    # result_dict is already a dictionary, no need for ast.literal_eval
+    fake_probability = result_dict.get('Fake Probability', 0.0)
+    logger.info(f"Debug: Parsed result_dict: {result_dict}, Extracted fake_probability: {fake_probability}")
+    return {"probabilities": np.array([fake_probability])} # Return as a numpy array with one element
+
+# New preprocess function for Gradio API
+def preprocess_gradio_api(image: Image.Image):
+    # The Gradio API expects a file path, so we need to save the PIL Image to a temporary file.
+    temp_file_path = "./temp_gradio_input.png"
+    image.save(temp_file_path)
+    return temp_file_path
+
+# New postprocess function for Gradio API (adapting postprocess_binary_output)
+def postprocess_gradio_api(gradio_output, class_names):
+    # gradio_output is expected to be a dictionary like {"probabilities": np.array([fake_prob])}
+    probabilities_array = None
+    if isinstance(gradio_output, dict) and "probabilities" in gradio_output:
+        probabilities_array = gradio_output["probabilities"]
+    elif isinstance(gradio_output, np.ndarray):
+        probabilities_array = gradio_output
+    else:
+        logger.warning(f"Unexpected output type for Gradio API post-processing: {type(gradio_output)}. Expected dict with 'probabilities' or numpy.ndarray.")
+        return {class_names[0]: 0.0, class_names[1]: 1.0}
+
+    logger.info(f"Debug: Probabilities array entering postprocess_gradio_api: {probabilities_array}, type: {type(probabilities_array)}, shape: {probabilities_array.shape}")
 
-def postprocess_pipeline(prediction, class_names):
-    # Assumes HuggingFace pipeline output
-    return {pred['label']: pred['score'] for pred in prediction}
+    if probabilities_array is None or probabilities_array.size == 0:
+        logger.warning("Probabilities array is None or empty after extracting from Gradio API output. Returning default scores.")
+        return {class_names[0]: 0.0, class_names[1]: 1.0}
 
-def postprocess_logits(outputs, class_names):
-    # Assumes model output with logits
-    logits = outputs.logits.cpu().numpy()[0]
-    probabilities = softmax(logits)
-    return {class_names[i]: probabilities[i] for i in range(len(class_names))}
+    # It should always be a single element array for fake probability
+    fake_prob = float(probabilities_array.item())
+    real_prob = 1.0 - fake_prob
+
+    return {class_names[0]: fake_prob, class_names[1]: real_prob}
 
 def register_model_with_metadata(model_id, model, preprocess, postprocess, class_names, display_name, contributor, model_path, architecture=None, dataset=None):
     entry = ModelEntry(model, preprocess, postprocess, class_names, display_name=display_name, contributor=contributor, model_path=model_path, architecture=architecture, dataset=dataset)
     MODEL_REGISTRY[model_id] = entry
 
-# Load and register models (copied from app_mcp.py)
-# image_processor_1 = AutoImageProcessor.from_pretrained(MODEL_PATHS["model_1"], use_fast=True)
-# model_1 = Swinv2ForImageClassification.from_pretrained(MODEL_PATHS["model_1"]).to(device)
-# clf_1 = pipeline(model=model_1, task="image-classification", image_processor=image_processor_1, device=device)
-# register_model_with_metadata(
-#     "model_1", clf_1, preprocess_resize_256, postprocess_pipeline, CLASS_NAMES["model_1"],
-#     display_name="SWIN1", contributor="haywoodsloan", model_path=MODEL_PATHS["model_1"],
-#     architecture="SwinV2", dataset="TBA"
-# )
 
-# --- ONNX Quantized Model Example ---
-ONNX_QUANTIZED_MODEL_PATH = "./models/model_1_quantized.onnx"
+def load_onnx_model_and_preprocessor(hf_model_id):
+    # model_dir = snapshot_download(repo_id=hf_model_id, local_dir_use_symlinks=False)
+    
+    # Create a unique local directory for each ONNX model
+    model_specific_dir = os.path.join("./models", hf_model_id.replace('/', '_'))
+    os.makedirs(model_specific_dir, exist_ok=True)
+
+    # Use hf_hub_download to get specific files into the model-specific directory
+    onnx_model_path = hf_hub_download(repo_id=hf_model_id, filename="model_quantized.onnx", subfolder="onnx", local_dir=model_specific_dir, local_dir_use_symlinks=False)
+
+    # Load preprocessor config
+    preprocessor_config = {}
+    try:
+        preprocessor_config_path = hf_hub_download(repo_id=hf_model_id, filename="preprocessor_config.json", local_dir=model_specific_dir, local_dir_use_symlinks=False)
+        with open(preprocessor_config_path, 'r') as f:
+            preprocessor_config = json.load(f)
+    except Exception as e:
+        logger.warning(f"Could not download or load preprocessor_config.json for {hf_model_id}: {e}")
+    
+    # Load model config for class names if available
+    model_config = {}
+    try:
+        model_config_path = hf_hub_download(repo_id=hf_model_id, filename="config.json", local_dir=model_specific_dir, local_dir_use_symlinks=False)
+        with open(model_config_path, 'r') as f:
+            model_config = json.load(f)
+    except Exception as e:
+        logger.warning(f"Could not download or load config.json for {hf_model_id}: {e}")
+
+    return onnxruntime.InferenceSession(onnx_model_path), preprocessor_config, model_config
+
 
-def preprocess_onnx_input(image: Image.Image):
-    # Preprocess image for ONNX model (e.g., for SwinV2, usually 256x256, normalized)
+# Cache for ONNX sessions and preprocessors
+_onnx_model_cache = {}
+
+def get_onnx_model_from_cache(hf_model_id):
+    if hf_model_id not in _onnx_model_cache:
+        logger.info(f"Loading ONNX model and preprocessor for {hf_model_id}...")
+        _onnx_model_cache[hf_model_id] = load_onnx_model_and_preprocessor(hf_model_id)
+    return _onnx_model_cache[hf_model_id]
+
+def preprocess_onnx_input(image: Image.Image, preprocessor_config: dict):
+    # Preprocess image for ONNX model based on preprocessor_config
     if image.mode != 'RGB':
         image = image.convert('RGB')
     
+    # Get image size and normalization values from preprocessor_config or use defaults
+    # Use 'size' for initial resize and 'crop_size' for center cropping
+    initial_resize_size = preprocessor_config.get('size', {'height': 224, 'width': 224})
+    crop_size = preprocessor_config.get('crop_size', initial_resize_size['height'])
+    mean = preprocessor_config.get('image_mean', [0.485, 0.456, 0.406])
+    std = preprocessor_config.get('image_std', [0.229, 0.224, 0.225])
+
     transform = transforms.Compose([
-        transforms.Resize((256, 256)),
+        transforms.Resize((initial_resize_size['height'], initial_resize_size['width'])),
+        transforms.CenterCrop(crop_size), # Apply center crop
         transforms.ToTensor(),
-        transforms.Normalize(mean=[0.485,0.456,0.406], std=[0.229,0.224,0.225]), # ImageNet normalization
+        transforms.Normalize(mean=mean, std=std),
     ])
     input_tensor = transform(image)
     # ONNX expects numpy array with batch dimension (1, C, H, W)
     return input_tensor.unsqueeze(0).cpu().numpy()
 
-def infer_onnx_model(preprocessed_image_np):
+def infer_onnx_model(hf_model_id, preprocessed_image_np, model_config: dict):
     try:
-        # Ensure the ONNX model exists before trying to load it
-        if not os.path.exists(ONNX_QUANTIZED_MODEL_PATH):
-            logger.error(f"ONNX quantized model not found at: {ONNX_QUANTIZED_MODEL_PATH}")
-            raise FileNotFoundError(f"ONNX quantized model not found at: {ONNX_QUANTIZED_MODEL_PATH}")
+        ort_session, _, _ = get_onnx_model_from_cache(hf_model_id)
+        
+        # Debug: Print expected input shape from ONNX model
+        for input_meta in ort_session.get_inputs():
+            logger.info(f"Debug: ONNX model expected input name: {input_meta.name}, shape: {input_meta.shape}, type: {input_meta.type}")
 
-        ort_session = onnxruntime.InferenceSession(ONNX_QUANTIZED_MODEL_PATH)
+        logger.info(f"Debug: preprocessed_image_np shape: {preprocessed_image_np.shape}")
         ort_inputs = {ort_session.get_inputs()[0].name: preprocessed_image_np}
         ort_outputs = ort_session.run(None, ort_inputs)
         
-        # Assuming the output is logits, apply softmax to get probabilities
         logits = ort_outputs[0]
-        probabilities = softmax(logits[0]) # Remove batch dim, apply softmax
+        logger.info(f"Debug: logits type: {type(logits)}, shape: {logits.shape}")
+        # If the model outputs a single logit (e.g., shape (1,)), use .item() to convert to scalar
+        # Otherwise, assume it's a batch of logits (e.g., shape (1, num_classes)) and take the first element (batch dim)
+        # The num_classes in config.json can be misleading; rely on actual output shape.
+        
+        # Apply softmax to the logits to get probabilities for the classes
+        # The softmax function in utils/utils.py now ensures a list of floats
+        probabilities = softmax(logits[0]) # Assuming logits[0] is the relevant output for a single prediction
+
         return {"logits": logits, "probabilities": probabilities}
 
     except Exception as e:
-        logger.error(f"Error during ONNX inference: {e}")
+        logger.error(f"Error during ONNX inference for {hf_model_id}: {e}")
         # Return a structure consistent with other model errors
         return {"logits": np.array([]), "probabilities": np.array([])}
 
-def postprocess_onnx_output(onnx_output, class_names):
+def postprocess_onnx_output(onnx_output, model_config):
+    # Get class names from model_config
+    # Prioritize id2label, then check num_classes, otherwise default
+    class_names_map = model_config.get('id2label')
+    if class_names_map:
+        class_names = [class_names_map[k] for k in sorted(class_names_map.keys())]
+    elif model_config.get('num_classes') == 1: # Handle models that output a single value (e.g., probability of 'Fake')
+        class_names = ['Fake', 'Real'] # Assume first class is 'Fake' and second 'Real'
+    else:
+        class_names = {0: 'Fake', 1: 'Real'} # Default to Fake/Real if not found or not 1 class
+        class_names = [class_names[i] for i in sorted(class_names.keys())]
+
     probabilities = onnx_output.get("probabilities")
-    if probabilities is not None and len(probabilities) == len(class_names):
-        return {class_names[i]: probabilities[i] for i in range(len(class_names))}
+    
+    if probabilities is not None:
+        if model_config.get('num_classes') == 1 and len(probabilities) == 2: # Special handling for single output models
+            # The single output is the probability of the 'Fake' class
+            fake_prob = float(probabilities[0])
+            real_prob = float(probabilities[1])
+            return {class_names[0]: fake_prob, class_names[1]: real_prob}
+        elif len(probabilities) == len(class_names):
+            return {class_names[i]: float(probabilities[i]) for i in range(len(class_names))}
+        else:
+            logger.warning("ONNX post-processing: Probabilities length mismatch with class names.")
+            return {name: 0.0 for name in class_names}
     else:
-        logger.warning("ONNX post-processing failed or class names mismatch.")
+        logger.warning("ONNX post-processing failed: 'probabilities' key not found in output.")
         return {name: 0.0 for name in class_names}
 
 # Register the ONNX quantized model
-register_model_with_metadata(
-    "model_1_onnx_quantized", 
-    infer_onnx_model, 
-    preprocess_onnx_input, 
-    postprocess_onnx_output, 
-    CLASS_NAMES["model_1"], # Assuming it uses the same class names as model_1
-    display_name="SWIN1", 
-    contributor="haywoodsloan", 
-    model_path=ONNX_QUANTIZED_MODEL_PATH,
-    architecture="SwinV2", 
-    dataset="TBA"
-)
-# --- End ONNX Quantized Model Example ---
-
-clf_2 = pipeline("image-classification", model=MODEL_PATHS["model_2"], device=device)
-register_model_with_metadata(
-    "model_2", clf_2, preprocess_resize_224, postprocess_pipeline, CLASS_NAMES["model_2"],
-    display_name="VIT2", contributor="Heem2", model_path=MODEL_PATHS["model_2"],
-    architecture="ViT", dataset="TBA"
-)
-
-feature_extractor_3 = AutoFeatureExtractor.from_pretrained(MODEL_PATHS["model_3"], device=device)
-model_3 = AutoModelForImageClassification.from_pretrained(MODEL_PATHS["model_3"]).to(device)
-def preprocess_256(image):
-    if image.mode != 'RGB':
-        image = image.convert('RGB')
-    return transforms.Resize((256, 256))(image)
-def postprocess_logits_model3(outputs, class_names):
-    logits = outputs.logits.cpu().numpy()[0]
-    probabilities = softmax(logits)
-    return {class_names[i]: probabilities[i] for i in range(len(class_names))}
-def model3_infer(image):
-    inputs = feature_extractor_3(image, return_tensors="pt").to(device)
-    with torch.no_grad():
-        outputs = model_3(**inputs)
-    return outputs
-register_model_with_metadata(
-    "model_3", model3_infer, preprocess_256, postprocess_logits_model3, CLASS_NAMES["model_3"],
-    display_name="SDXL3", contributor="Organika", model_path=MODEL_PATHS["model_3"],
-    architecture="VIT", dataset="SDXL"
-)
-
-feature_extractor_4 = AutoFeatureExtractor.from_pretrained(MODEL_PATHS["model_4"], device=device)
-model_4 = AutoModelForImageClassification.from_pretrained(MODEL_PATHS["model_4"]).to(device)
-def model4_infer(image):
-    inputs = feature_extractor_4(image, return_tensors="pt").to(device)
-    with torch.no_grad():
-        outputs = model_4(**inputs)
-    return outputs
-def postprocess_logits_model4(outputs, class_names):
-    logits = outputs.logits.cpu().numpy()[0]
-    probabilities = softmax(logits)
-    return {class_names[i]: probabilities[i] for i in range(len(class_names))}
-register_model_with_metadata(
-    "model_4", model4_infer, preprocess_256, postprocess_logits_model4, CLASS_NAMES["model_4"],
-    display_name="XLFLUX4", contributor="cmckinle", model_path=MODEL_PATHS["model_4"],
-    architecture="VIT", dataset="SDXL, FLUX"
-)
-
-clf_5 = pipeline("image-classification", model=MODEL_PATHS["model_5"], device=device)
-register_model_with_metadata(
-    "model_5", clf_5, preprocess_resize_224, postprocess_pipeline, CLASS_NAMES["model_5"],
-    display_name="VIT5", contributor="prithivMLmods", model_path=MODEL_PATHS["model_5"],
-    architecture="VIT", dataset="TBA"
-)
-
-image_processor_6 = AutoImageProcessor.from_pretrained(MODEL_PATHS["model_6"], use_fast=True)
-model_6 = SwinForImageClassification.from_pretrained(MODEL_PATHS["model_6"]).to(device)
-clf_6 = pipeline(model=model_6, task="image-classification", image_processor=image_processor_6, device=device)
-register_model_with_metadata(
-    "model_6", clf_6, preprocess_resize_224, postprocess_pipeline, CLASS_NAMES["model_6"],
-    display_name="SWIN6", contributor="ideepankarsharma2003", model_path=MODEL_PATHS["model_6"],
-    architecture="SWINv1", dataset="SDXL, Midjourney"
-)
+# Dummy entry for ONNX model to be loaded dynamically
+# We will now register a 'wrapper' that handles dynamic loading
+
+class ONNXModelWrapper:
+    def __init__(self, hf_model_id):
+        self.hf_model_id = hf_model_id
+        self._session = None
+        self._preprocessor_config = None
+        self._model_config = None
+
+    def load(self):
+        if self._session is None:
+            self._session, self._preprocessor_config, self._model_config = get_onnx_model_from_cache(self.hf_model_id)
+            logger.info(f"ONNX model {self.hf_model_id} loaded into wrapper.")
+
+    def __call__(self, image_np):
+        self.load() # Ensure model is loaded on first call
+        # Pass model_config to infer_onnx_model
+        return infer_onnx_model(self.hf_model_id, image_np, self._model_config)
+
+    def preprocess(self, image: Image.Image):
+        self.load()
+        return preprocess_onnx_input(image, self._preprocessor_config)
+
+    def postprocess(self, onnx_output: dict, class_names_from_registry: list): # class_names_from_registry is ignored
+        self.load()
+        return postprocess_onnx_output(onnx_output, self._model_config)
+
+# Consolidate all model loading and registration
+for model_key, hf_model_path in MODEL_PATHS.items():
+    logger.debug(f"Attempting to register model: {model_key} with path: {hf_model_path}")
+    model_num = model_key.replace("model_", "").upper()
+    contributor = "Unknown"
+    architecture = "Unknown"
+    dataset = "TBA"
+
+    current_class_names = CLASS_NAMES.get(model_key, [])
+
+    # Logic for ONNX models (1, 2, 3, 5, 6, 7)
+    if "ONNX" in hf_model_path:
+        logger.debug(f"Model {model_key} identified as ONNX.")
+        logger.info(f"Registering ONNX model: {model_key} from {hf_model_path}")
+        onnx_wrapper_instance = ONNXModelWrapper(hf_model_path)
+
+        # Attempt to derive contributor, architecture, dataset based on model_key
+        if model_key == "model_1":
+            contributor = "haywoodsloan"
+            architecture = "SwinV2"
+            dataset = "DeepFakeDetection"
+        elif model_key == "model_2":
+            contributor = "Heem2"
+            architecture = "ViT"
+            dataset = "DeepFakeDetection"
+        elif model_key == "model_3":
+            contributor = "Organika"
+            architecture = "VIT"
+            dataset = "SDXL"
+        elif model_key == "model_5":
+            contributor = "prithivMLmods"
+            architecture = "VIT"
+        elif model_key == "model_6":
+            contributor = "ideepankarsharma2003"
+            architecture = "SWINv1"
+            dataset = "SDXL, Midjourney"
+        elif model_key == "model_7":
+            contributor = "date3k2"
+            architecture = "VIT"
+
+        display_name_parts = [model_num]
+        if architecture and architecture not in ["Unknown"]:
+            display_name_parts.append(architecture)
+        if dataset and dataset not in ["TBA"]:
+            display_name_parts.append(dataset)
+        display_name = "-".join(display_name_parts)
+        display_name += "_ONNX" # Always append _ONNX for ONNX models
+
+        register_model_with_metadata(
+            model_id=model_key,
+            model=onnx_wrapper_instance, # The callable wrapper for the ONNX model
+            preprocess=onnx_wrapper_instance.preprocess,
+            postprocess=onnx_wrapper_instance.postprocess,
+            class_names=current_class_names, # Initial class names; will be overridden by model_config if available
+            display_name=display_name,
+            contributor=contributor,
+            model_path=hf_model_path,
+            architecture=architecture,
+            dataset=dataset
+        )
+    # Logic for Gradio API model (model_8)
+    elif model_key == "model_8":
+        logger.debug(f"Model {model_key} identified as Gradio API.")
+        logger.info(f"Registering Gradio API model: {model_key} from {hf_model_path}")
+        contributor = "aiwithoutborders-xyz"
+        architecture = "ViT"
+        dataset = "DeepfakeDetection"
+
+        display_name_parts = [model_num]
+        if architecture and architecture not in ["Unknown"]:
+            display_name_parts.append(architecture)
+        if dataset and dataset not in ["TBA"]:
+            display_name_parts.append(dataset)
+        display_name = "-".join(display_name_parts)
+
+        register_model_with_metadata(
+            model_id=model_key,
+            model=infer_gradio_api,
+            preprocess=preprocess_gradio_api,
+            postprocess=postprocess_gradio_api,
+            class_names=current_class_names,
+            display_name=display_name,
+            contributor=contributor,
+            model_path=hf_model_path,
+            architecture=architecture,
+            dataset=dataset
+        )
+    # Logic for PyTorch/Hugging Face pipeline models (currently only model_4)
+    elif model_key == "model_4": # Explicitly handle model_4
+        logger.debug(f"Model {model_key} identified as PyTorch/HuggingFace pipeline.")
+        logger.info(f"Registering HuggingFace pipeline/AutoModel: {model_key} from {hf_model_path}")
+        contributor = "cmckinle"
+        architecture = "VIT"
+        dataset = "SDXL, FLUX"
+
+        display_name_parts = [model_num]
+        if architecture and architecture not in ["Unknown"]:
+            display_name_parts.append(architecture)
+        if dataset and dataset not in ["TBA"]:
+            display_name_parts.append(dataset)
+        display_name = "-".join(display_name_parts)
+
+        current_processor = AutoFeatureExtractor.from_pretrained(hf_model_path, device=device)
+        model_instance = AutoModelForImageClassification.from_pretrained(hf_model_path).to(device)
+
+        preprocess_func = preprocess_resize_256
+        postprocess_func = postprocess_logits
+
+        def custom_infer(image, processor_local=current_processor, model_local=model_instance):
+            inputs = processor_local(image, return_tensors="pt").to(device)
+            with torch.no_grad():
+                outputs = model_local(**inputs)
+            return outputs
+        model_instance = custom_infer
+
+        register_model_with_metadata(
+            model_id=model_key,
+            model=model_instance,
+            preprocess=preprocess_func,
+            postprocess=postprocess_func,
+            class_names=current_class_names,
+            display_name=display_name,
+            contributor=contributor,
+            model_path=hf_model_path,
+            architecture=architecture,
+            dataset=dataset
+        )
+    else: # Fallback for any unhandled models (shouldn't happen if MODEL_PATHS is fully covered)
+        logger.warning(f"Could not automatically load and register model: {model_key} from {hf_model_path}. No matching registration logic found.")
 
-image_processor_7 = AutoImageProcessor.from_pretrained(MODEL_PATHS["model_7"], use_fast=True)
-model_7 = AutoModelForImageClassification.from_pretrained(MODEL_PATHS["model_7"]).to(device)
-clf_7 = pipeline(model=model_7, task="image-classification", image_processor=image_processor_7, device=device)
-register_model_with_metadata(
-    "model_7", clf_7, preprocess_resize_224, postprocess_pipeline, CLASS_NAMES["model_7"],
-    display_name="VIT7", contributor="date3k2", model_path=MODEL_PATHS["model_7"],
-    architecture="VIT", dataset="TBA"
-)
 
-# def postprocess_simple_prediction(result, class_names):
-#     scores = {name: 0.0 for name in class_names}
-#     fake_prob = result.get("Fake Probability")
-#     if fake_prob is not None:
-#         # Assume class_names = ["AI", "REAL"]
-#         scores["AI"] = float(fake_prob)
-#         scores["REAL"] = 1.0 - float(fake_prob)
-#     return scores
-
-# def simple_prediction(img):
-#     client = Client("aiwithoutborders-xyz/OpenSight-Community-Forensics-Preview")
-#     client.view_api()
-#     print(type(img))
-#     result = client.predict(
-#            handle_file(img),
-#            api_name="simple_predict"
-#     )
-#     return result
-
-
-# register_model_with_metadata(
-#     model_id="simple_prediction",
-#     model=simple_prediction,
-#     preprocess=None,
-#     postprocess=postprocess_simple_prediction,
-#     class_names=["AI", "REAL"], 
-#     display_name="Community Forensics",
-#     contributor="Jeongsoo Park",
-#     model_path="aiwithoutborders-xyz/CommunityForensics-DeepfakeDet-ViT",
-#     architecture="ViT", dataset="GOAT"
-# )
 def infer(image: Image.Image, model_id: str, confidence_threshold: float = 0.75) -> dict:
     """Predict using a specific model.
 
@@ -297,8 +434,14 @@ def infer(image: Image.Image, model_id: str, confidence_threshold: float = 0.75)
     try:
         result = entry.model(img)
         scores = entry.postprocess(result, entry.class_names)
-        ai_score = float(scores.get(entry.class_names[0], 0.0))
-        real_score = float(scores.get(entry.class_names[1], 0.0))
+        
+        def _to_float_scalar(value):
+            if isinstance(value, np.ndarray):
+                return float(value.item()) # Convert numpy array scalar to Python float
+            return float(value) # Already a Python scalar or convertible type
+
+        ai_score = _to_float_scalar(scores.get(entry.class_names[0], 0.0))
+        real_score = _to_float_scalar(scores.get(entry.class_names[1], 0.0))
         label = "AI" if ai_score >= confidence_threshold else ("REAL" if real_score >= confidence_threshold else "UNCERTAIN")
         return {
             "Model": entry.display_name,
@@ -371,29 +514,97 @@ def full_prediction(img, confidence_threshold, rotate_degrees, noise_level, shar
     results = []
     table_rows = []
 
+    # Initialize lists for forensic outputs, starting with the original augmented image
+    cleaned_forensics_images = []
+    forensic_output_descriptions = []
+
+    # Always add the original augmented image first for forensic display
+    if isinstance(img_pil, Image.Image):
+        cleaned_forensics_images.append(img_pil)
+        forensic_output_descriptions.append(f"Original augmented image (PIL): {img_pil.width}x{img_pil.height}")
+    elif isinstance(img_pil, np.ndarray):
+        try:
+            pil_img_from_np = Image.fromarray(img_pil)
+            cleaned_forensics_images.append(pil_img_from_np)
+            forensic_output_descriptions.append(f"Original augmented image (numpy converted to PIL): {pil_img_from_np.width}x{pil_img_from_np.height}")
+        except Exception as e:
+            logger.warning(f"Could not convert original numpy image to PIL for gallery: {e}")
+    
+    # Yield initial state with augmented image and empty model predictions
+    yield img_pil, cleaned_forensics_images, table_rows, "[]", "<div style='font-size: 2.2em; font-weight: bold;padding: 10px;'>Consensus: <span style='color:orange'>UNCERTAIN</span></div>"
+
+
     # Stream results as each model finishes
     for model_id in MODEL_REGISTRY:
         model_start = time.time()
         result = infer(img_pil, model_id, confidence_threshold)
         model_end = time.time()
+
+        # Helper to ensure values are Python floats, handling numpy scalars
+        def _ensure_float_scalar(value):
+            if isinstance(value, np.ndarray):
+                return float(value.item()) # Convert numpy array scalar to Python float
+            return float(value) # Already a Python scalar or convertible type
+
+        ai_score_val = _ensure_float_scalar(result.get("AI Score", 0.0))
+        real_score_val = _ensure_float_val = _ensure_float_scalar(result.get("Real Score", 0.0))
+
         monitor_agent.monitor_prediction(
             model_id,
             result["Label"],
-            max(result.get("AI Score", 0.0), result.get("Real Score", 0.0)),
+            max(ai_score_val, real_score_val),
             model_end - model_start
         )
         model_predictions_raw[model_id] = result
-        confidence_scores[model_id] = max(result.get("AI Score", 0.0), result.get("Real Score", 0.0))
+        confidence_scores[model_id] = max(ai_score_val, real_score_val)
         results.append(result)
         table_rows.append([
             result.get("Model", ""),
             result.get("Contributor", ""),
-            round(result.get("AI Score", 0.0), 3) if result.get("AI Score") is not None else 0.0,
-            round(result.get("Real Score", 0.0), 3) if result.get("Real Score") is not None else 0.0,
+            round(ai_score_val, 5),
+            round(real_score_val, 5),
             result.get("Label", "Error")
         ])
         # Yield partial results: only update the table, others are None
-        yield None, None, table_rows, None, None
+        yield None, cleaned_forensics_images, table_rows, None, None # Keep cleaned_forensics_images as is (only augmented image for now)
+
+    # Multi-threaded forensic processing
+    def _run_forensic_task(task_func, img_input, description, **kwargs):
+        try:
+            result_img = task_func(img_input, **kwargs)
+            return result_img, description
+        except Exception as e:
+            logger.error(f"Error processing forensic task {task_func.__name__}: {e}")
+            return None, f"Error processing {description}: {str(e)}"
+
+    with concurrent.futures.ThreadPoolExecutor() as executor:
+        future_ela1 = executor.submit(_run_forensic_task, ELA, img_np_og, "ELA analysis (Pass 1): Grayscale error map, quality 75.", quality=75, scale=50, contrast=20, linear=False, grayscale=True)
+        future_ela2 = executor.submit(_run_forensic_task, ELA, img_np_og, "ELA analysis (Pass 2): Grayscale error map, quality 75, enhanced contrast.", quality=75, scale=75, contrast=25, linear=False, grayscale=True)
+        future_ela3 = executor.submit(_run_forensic_task, ELA, img_np_og, "ELA analysis (Pass 3): Color error map, quality 75, enhanced contrast.", quality=75, scale=75, contrast=25, linear=False, grayscale=False)
+        future_gradient1 = executor.submit(_run_forensic_task, gradient_processing, img_np_og, "Gradient processing: Highlights edges and transitions.")
+        future_gradient2 = executor.submit(_run_forensic_task, gradient_processing, img_np_og, "Gradient processing: Int=45, Equalize=True", intensity=45, equalize=True)
+        future_minmax1 = executor.submit(_run_forensic_task, minmax_process, img_np_og, "MinMax processing: Deviations in local pixel values.")
+        future_minmax2 = executor.submit(_run_forensic_task, minmax_process, img_np_og, "MinMax processing (Radius=6): Deviations in local pixel values.", radius=6)
+
+        forensic_futures = [future_ela1, future_ela2, future_ela3, future_gradient1, future_gradient2, future_minmax1, future_minmax2]
+        
+        for future in concurrent.futures.as_completed(forensic_futures):
+            processed_img, description = future.result()
+            if processed_img is not None:
+                if isinstance(processed_img, Image.Image):
+                    cleaned_forensics_images.append(processed_img)
+                elif isinstance(processed_img, np.ndarray):
+                    try:
+                        cleaned_forensics_images.append(Image.fromarray(processed_img))
+                    except Exception as e:
+                        logger.warning(f"Could not convert numpy array to PIL Image for gallery: {e}")
+                else:
+                    logger.warning(f"Unexpected type in processed_img from {description}: {type(processed_img)}. Skipping.")
+                
+                forensic_output_descriptions.append(description) # Keep track of descriptions for anomaly agent
+                
+                # Yield partial results: update gallery
+                yield None, cleaned_forensics_images, table_rows, None, None
 
     # After all models, compute the rest as before
     image_data_for_context = {
@@ -401,6 +612,17 @@ def full_prediction(img, confidence_threshold, rotate_degrees, noise_level, shar
         "height": img.height,
         "mode": img.mode,
     }
+    forensic_output_descriptions = [
+        f"Original augmented image (PIL): {img_pil.width}x{img_pil.height}",
+        "ELA analysis (Pass 1): Grayscale error map, quality 75.",
+        "ELA analysis (Pass 2): Grayscale error map, quality 75, enhanced contrast.",
+        "ELA analysis (Pass 3): Color error map, quality 75, enhanced contrast.",
+        "Gradient processing: Highlights edges and transitions.",
+        "Gradient processing: Int=45, Equalize=True",
+        "MinMax processing: Deviations in local pixel values.",
+        "MinMax processing (Radius=6): Deviations in local pixel values.",
+        # "Bit Plane extractor: Visualization of individual bit planes from different color channels."
+    ]
     detected_context_tags = context_agent.infer_context_tags(image_data_for_context, model_predictions_raw)
     logger.info(f"Detected context tags: {detected_context_tags}")
     adjusted_weights = weight_manager.adjust_weights(model_predictions_raw, confidence_scores, context_tags=detected_context_tags)
@@ -417,26 +639,26 @@ def full_prediction(img, confidence_threshold, rotate_degrees, noise_level, shar
     elif weighted_predictions["REAL"] > weighted_predictions["AI"] and weighted_predictions["REAL"] > weighted_predictions["UNCERTAIN"]:
         final_prediction_label = "REAL"
     optimization_agent.analyze_performance(final_prediction_label, None)
-    gradient_image = gradient_processing(img_np_og)
-    gradient_image2 = gradient_processing(img_np_og, intensity=45, equalize=True)
-    minmax_image = minmax_process(img_np_og)
-    minmax_image2 = minmax_process(img_np_og, radius=6)
-    # bitplane_image = bit_plane_extractor(img_pil)
-    ela1 = ELA(img_np_og, quality=75, scale=50, contrast=20, linear=False, grayscale=True)
-    ela2 = ELA(img_np_og, quality=75, scale=75, contrast=25, linear=False, grayscale=True)
-    ela3 = ELA(img_np_og, quality=75, scale=75, contrast=25, linear=False, grayscale=False)
-    forensics_images = [img_pil, ela1, ela2, ela3, gradient_image, gradient_image2, minmax_image, minmax_image2]
-    forensic_output_descriptions = [
-        f"Original augmented image (PIL): {img_pil.width}x{img_pil.height}",
-        "ELA analysis (Pass 1): Grayscale error map, quality 75.",
-        "ELA analysis (Pass 2): Grayscale error map, quality 75, enhanced contrast.",
-        "ELA analysis (Pass 3): Color error map, quality 75, enhanced contrast.",
-        "Gradient processing: Highlights edges and transitions.",
-        "Gradient processing: Int=45, Equalize=True",
-        "MinMax processing: Deviations in local pixel values.",
-        "MinMax processing (Radius=6): Deviations in local pixel values.",
-        # "Bit Plane extractor: Visualization of individual bit planes from different color channels."
-    ]
+    # gradient_image = gradient_processing(img_np_og)
+    # gradient_image2 = gradient_processing(img_np_og, intensity=45, equalize=True)
+    # minmax_image = minmax_process(img_np_og)
+    # minmax_image2 = minmax_process(img_np_og, radius=6)
+    # # bitplane_image = bit_plane_extractor(img_pil)
+    # ela1 = ELA(img_np_og, quality=75, scale=50, contrast=20, linear=False, grayscale=True)
+    # ela2 = ELA(img_np_og, quality=75, scale=75, contrast=25, linear=False, grayscale=True)
+    # ela3 = ELA(img_np_og, quality=75, scale=75, contrast=25, linear=False, grayscale=False)
+    # forensics_images = [img_pil, ela1, ela2, ela3, gradient_image, gradient_image2, minmax_image, minmax_image2]
+    # forensic_output_descriptions = [
+    #     f"Original augmented image (PIL): {img_pil.width}x{img_pil.height}",
+    #     "ELA analysis (Pass 1): Grayscale error map, quality 75.",
+    #     "ELA analysis (Pass 2): Grayscale error map, quality 75, enhanced contrast.",
+    #     "ELA analysis (Pass 3): Color error map, quality 75, enhanced contrast.",
+    #     "Gradient processing: Highlights edges and transitions.",
+    #     "Gradient processing: Int=45, Equalize=True",
+    #     "MinMax processing: Deviations in local pixel values.",
+    #     "MinMax processing (Radius=6): Deviations in local pixel values.",
+    #     # "Bit Plane extractor: Visualization of individual bit planes from different color channels."
+    # ]
     anomaly_detection_results = anomaly_agent.analyze_forensic_outputs(forensic_output_descriptions)
     logger.info(f"Forensic anomaly detection: {anomaly_detection_results['summary']}")
     consensus_html = f"<div style='font-size: 2.2em; font-weight: bold;padding: 10px;'>Consensus: <span style='color:{'red' if final_prediction_label == 'AI' else ('green' if final_prediction_label == 'REAL' else 'orange')}'>{final_prediction_label}</span></div>"
@@ -474,21 +696,11 @@ def full_prediction(img, confidence_threshold, rotate_degrees, noise_level, shar
         inference_params=inference_params,
         model_predictions=results,
         ensemble_output=ensemble_output_data,
-        forensic_images=forensics_images,
+        forensic_images=cleaned_forensics_images, # Use the incrementally built list
         agent_monitoring_data=agent_monitoring_data_log,
         human_feedback=None
     )
-    cleaned_forensics_images = []
-    for f_img in forensics_images:
-        if isinstance(f_img, Image.Image):
-            cleaned_forensics_images.append(f_img)
-        elif isinstance(f_img, np.ndarray):
-            try:
-                cleaned_forensics_images.append(Image.fromarray(f_img))
-            except Exception as e:
-                logger.warning(f"Could not convert numpy array to PIL Image for gallery: {e}")
-        else:
-            logger.warning(f"Unexpected type in forensic_images: {type(f_img)}. Skipping.")
+
     logger.info(f"Cleaned forensic images types: {[type(img) for img in cleaned_forensics_images]}")
     for i, res_dict in enumerate(results):
         for key in ["AI Score", "Real Score"]:
@@ -709,7 +921,7 @@ demo = gr.TabbedInterface(
 
 )
 footerMD = """
-### ⚠️ ENSEMBLE TEAM IN TRAINING ⚠️ \n\n
+## ⚠️ ENSEMBLE TEAM IN TRAINING ⚠️ \n\n
 
 1. **DISCLAIMER: METADATA AS WELL AS MEDIA SUBMITTED TO THIS SPACE MAY BE VIEWED AND SELECTED FOR FUTURE DATASETS, PLEASE DO NOT SUBMIT PERSONAL CONTENT. FOR UNTRACKED, PRIVATE USE OF THE MODELS YOU MAY STILL USE [THE ORIGINAL SPACE HERE](https://huggingface.co/spaces/aiwithoutborders-xyz/OpenSight-Deepfake-Detection-Models-Playground), SOTA MODEL INCLUDED.**
 2. **UPDATE 6-13-25**: APOLOGIES FOR THE CONFUSION, WE ARE WORKING TO REVERT THE ORIGINAL REPO BACK TO ITS NON-DATA COLLECTION STATE -- ONLY THE "SIMPLE PREDICTION" ENDPOINT IS CURRENTLY 100% PRIVATE. PLEASE STAY TUNED AS WE FIGURE OUT A SOLUTION FOR THE ENSEMBLE + AGENT TEAM ENDPOINT. IT CAN GET RESOURCE INTENSIVE TO RUN A FULL PREDICTION. ALTERNATIVELY, WE **ENCOURAGE** ANYONE TO FORK AND CONTRIBUTE TO THE PROJECT.
@@ -718,7 +930,7 @@ footerMD = """
 TO SUMMARIZE: DATASET COLLECTION WILL CONTINUE FOR OUR NOVEL ENSEMBLE-TEAM PREDICTION PIPELINE UNTIL WE CAN GET THINGS SORTED OUT. FOR THOSE THAT WISH TO OPT-OUT, WE OFFER THE SIMPLE, BUT [MOST POWERFUL DETECTION MODEL HERE.](https://huggingface.co/spaces/aiwithoutborders-xyz/OpenSight-Community-Forensics-Preview)
 
 """
-footer = gr.Markdown("", elem_classes="footer")
+footer = gr.Markdown(footerMD, elem_classes="footer")
 
 with gr.Blocks() as app:
     demo.render()

utils/utils.py

@@ -1,3 +1,57 @@
+def preprocess_resize_256(image):
+    if image.mode != 'RGB':
+        image = image.convert('RGB')
+    return transforms.Resize((256, 256))(image)
+
+def preprocess_resize_224(image):
+    if image.mode != 'RGB':
+        image = image.convert('RGB')
+    return transforms.Resize((224, 224))(image)
+
+def postprocess_pipeline(prediction, class_names):
+    # Assumes HuggingFace pipeline output
+    return {pred['label']: float(pred['score']) for pred in prediction}
+
+def postprocess_logits(outputs, class_names):
+    # Assumes model output with logits
+    logits = outputs.logits.cpu().numpy()[0]
+    probabilities = softmax(logits)
+    return {class_names[i]: probabilities[i] for i in range(len(class_names))}
+
+def postprocess_binary_output(output, class_names):
+    # output can be a dictionary {"probabilities": numpy_array} or directly a numpy_array
+    import logging
+    logger = logging.getLogger(__name__)
+    probabilities_array = None
+    if isinstance(output, dict) and "probabilities" in output:
+        probabilities_array = output["probabilities"]
+    elif isinstance(output, np.ndarray):
+        probabilities_array = output
+    else:
+        logger.warning(f"Unexpected output type for binary post-processing: {type(output)}. Expected dict with 'probabilities' or numpy.ndarray.")
+        return {class_names[0]: 0.0, class_names[1]: 1.0}
+
+    logger.info(f"Debug: Probabilities array entering postprocess_binary_output: {probabilities_array}, type: {type(probabilities_array)}, shape: {getattr(probabilities_array, 'shape', None)}")
+
+    if probabilities_array is None:
+        logger.warning("Probabilities array is None after extracting from output. Returning default scores.")
+        return {class_names[0]: 0.0, class_names[1]: 1.0}
+
+    if probabilities_array.size == 1:
+        fake_prob = float(probabilities_array.item())
+    elif probabilities_array.size == 2:
+        fake_prob = float(probabilities_array[0])
+    else:
+        logger.warning(f"Unexpected probabilities array shape for binary post-processing: {probabilities_array.shape}. Expected size 1 or 2.")
+        return {class_names[0]: 0.0, class_names[1]: 1.0}
+
+    real_prob = 1.0 - fake_prob # Ensure Fake and Real sum to 1
+    return {class_names[0]: fake_prob, class_names[1]: real_prob}
+
+def to_float_scalar(value):
+    if isinstance(value, np.ndarray):
+        return float(value.item()) # Convert numpy array scalar to Python float
+    return float(value) # Already a Python scalar or convertible type
 import numpy as np
 import io
 from PIL import Image, ImageFilter, ImageChops