import gradio as gr from gradio_client import Client, handle_file from PIL import Image, ImageFilter import numpy as np import os import time 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 from forensics.minmax import minmax_process from forensics.ela import ELA from forensics.wavelet import noise_estimation from forensics.bitplane import bit_plane_extractor from utils.hf_logger import log_inference_data from utils.load import load_image from agents.ensemble_team import EnsembleMonitorAgent, WeightOptimizationAgent, SystemHealthAgent from agents.smart_agents import ContextualIntelligenceAgent, ForensicAnomalyDetectionAgent 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 logging.basicConfig(level=logging.INFO) logger = logging.getLogger(__name__) os.environ['HF_HUB_CACHE'] = './models' # --- Gradio Log Handler --- class GradioLogHandler(logging.Handler): def __init__(self, log_queue): super().__init__() self.log_queue = log_queue self.setFormatter(logging.Formatter('%(asctime)s - %(name)s - %(levelname)s - %(message)s')) def emit(self, record): self.log_queue.append(self.format(record)) log_queue = collections.deque(maxlen=1000) # Store last 1000 log messages gradio_handler = GradioLogHandler(log_queue) # Set root logger level to DEBUG to capture all messages from agents logging.getLogger().setLevel(logging.INFO) logging.getLogger().addHandler(gradio_handler) # --- End Gradio Log Handler --- LOCAL_LOG_DIR = "./hf_inference_logs" HF_DATASET_NAME="aiwithoutborders-xyz/degentic_rd0" load_dotenv() # Custom JSON Encoder to handle numpy types class NumpyEncoder(json.JSONEncoder): def default(self, obj): if isinstance(obj, np.float32): return float(obj) return json.JSONEncoder.default(self, obj) # Ensure using GPU if available 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_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" } CLASS_NAMES = { "model_1": ['artificial', 'real'], "model_2": ['AI Image', 'Real Image'], "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 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 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 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] 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((initial_resize_size['height'], initial_resize_size['width'])), transforms.CenterCrop(crop_size), # Apply center crop transforms.ToTensor(), 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(hf_model_id, preprocessed_image_np, model_config: dict): 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}") 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) 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 return {"logits": logits, "probabilities": probabilities} except Exception as 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, 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: 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: 'probabilities' key not found in output.") 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.") def infer(image: Image.Image, model_id: str, confidence_threshold: float = 0.75) -> dict: """Predict using a specific model. Args: image (Image.Image): The input image to classify. model_id (str): The ID of the model to use for classification. confidence_threshold (float, optional): The confidence threshold for classification. Defaults to 0.75. Returns: dict: A dictionary containing the model details, classification scores, and label. """ entry = MODEL_REGISTRY[model_id] img = entry.preprocess(image) if entry.preprocess else image 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)) label = "AI" if ai_score >= confidence_threshold else ("REAL" if real_score >= confidence_threshold else "UNCERTAIN") return { "Model": entry.display_name, "Contributor": entry.contributor, "HF Model Path": entry.model_path, "AI Score": ai_score, "Real Score": real_score, "Label": label } except Exception as e: return { "Model": entry.display_name, "Contributor": entry.contributor, "HF Model Path": entry.model_path, "AI Score": 0.0, "Real Score": 0.0, "Label": f"Error: {str(e)}" } def full_prediction(img, confidence_threshold, rotate_degrees, noise_level, sharpen_strength): """Full prediction run, with a team of ensembles and agents. Args: img (url: str, Image.Image, np.ndarray): The input image to classify. confidence_threshold (float, optional): The confidence threshold for classification. Defaults to 0.75. rotate_degrees (int, optional): The degrees to rotate the image. noise_level (int, optional): The noise level to use. sharpen_strength (int, optional): The sharpen strength to use. Returns: dict: A dictionary containing the model details, classification scores, and label. """ # Ensure img is a PIL Image object if img is None: raise gr.Error("No image provided. Please upload an image to analyze.") # Handle filepath conversion if needed if isinstance(img, str): try: img = load_image(img) except Exception as e: logger.error(f"Error loading image from path: {e}") raise gr.Error(f"Could not load image from the provided path. Error: {str(e)}") if not isinstance(img, Image.Image): try: img = Image.fromarray(img) except Exception as e: logger.error(f"Error converting input image to PIL: {e}") raise gr.Error("Input image could not be converted to a valid image format. Please try another image.") # Ensure image is in RGB format for consistent processing if img.mode != 'RGB': img = img.convert('RGB') monitor_agent = EnsembleMonitorAgent() weight_manager = ModelWeightManager(strongest_model_id="simple_prediction") optimization_agent = WeightOptimizationAgent(weight_manager) health_agent = SystemHealthAgent() context_agent = ContextualIntelligenceAgent() anomaly_agent = ForensicAnomalyDetectionAgent() health_agent.monitor_system_health() if rotate_degrees or noise_level or sharpen_strength: img_pil, _ = augment_image(img, ["rotate", "add_noise", "sharpen"], rotate_degrees, noise_level, sharpen_strength) else: img_pil = img img_np_og = np.array(img) model_predictions_raw = {} confidence_scores = {} 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, "[]", "
Consensus: UNCERTAIN
" # 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), model_end - model_start ) model_predictions_raw[model_id] = result confidence_scores[model_id] = max(ai_score_val, real_score_val) results.append(result) table_rows.append([ result.get("Model", ""), result.get("Contributor", ""), 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, 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 = { "width": img.width, "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) weighted_predictions = {"AI": 0.0, "REAL": 0.0, "UNCERTAIN": 0.0} for model_id, prediction in model_predictions_raw.items(): prediction_label = prediction.get("Label") if prediction_label in weighted_predictions: weighted_predictions[prediction_label] += adjusted_weights[model_id] else: logger.warning(f"Unexpected prediction label '{prediction_label}' from model '{model_id}'. Skipping its weight in consensus.") final_prediction_label = "UNCERTAIN" if weighted_predictions["AI"] > weighted_predictions["REAL"] and weighted_predictions["AI"] > weighted_predictions["UNCERTAIN"]: final_prediction_label = "AI" 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." # ] anomaly_detection_results = anomaly_agent.analyze_forensic_outputs(forensic_output_descriptions) logger.info(f"Forensic anomaly detection: {anomaly_detection_results['summary']}") consensus_html = f"
Consensus: {final_prediction_label}
" inference_params = { "confidence_threshold": confidence_threshold, "rotate_degrees": rotate_degrees, "noise_level": noise_level, "sharpen_strength": sharpen_strength, "detected_context_tags": detected_context_tags } ensemble_output_data = { "final_prediction_label": final_prediction_label, "weighted_predictions": weighted_predictions, "adjusted_weights": adjusted_weights } agent_monitoring_data_log = { "ensemble_monitor": { "alerts": monitor_agent.alerts, "performance_metrics": monitor_agent.performance_metrics }, "weight_optimization": { "prediction_history_length": len(optimization_agent.prediction_history), }, "system_health": { "memory_usage": health_agent.health_metrics["memory_usage"], "gpu_utilization": health_agent.health_metrics["gpu_utilization"] }, "context_intelligence": { "detected_context_tags": detected_context_tags }, "forensic_anomaly_detection": anomaly_detection_results } log_inference_data( original_image=img, inference_params=inference_params, model_predictions=results, ensemble_output=ensemble_output_data, forensic_images=cleaned_forensics_images, # Use the incrementally built list agent_monitoring_data=agent_monitoring_data_log, human_feedback=None ) 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"]: value = res_dict.get(key) if isinstance(value, np.float32): res_dict[key] = float(value) logger.info(f"Converted {key} for result {i} from numpy.float32 to float.") json_results = json.dumps(results, cls=NumpyEncoder) yield img_pil, cleaned_forensics_images, table_rows, json_results, consensus_html detection_model_eval_playground = gr.Interface( fn=full_prediction, inputs=[ gr.Image(label="Upload Image to Analyze", sources=['upload', 'webcam'], type='filepath'), gr.Slider(0.0, 1.0, value=0.7, step=0.05, label="Confidence Threshold"), gr.Slider(0, 45, value=0, step=1, label="Rotate Degrees", visible=False), gr.Slider(0, 50, value=0, step=1, label="Noise Level", visible=False), gr.Slider(0, 50, value=0, step=1, label="Sharpen Strength", visible=False) ], outputs=[ gr.Image(label="Processed Image", visible=False), gr.Gallery(label="Post Processed Images", visible=True, columns=[4], rows=[2], container=False, height="auto", object_fit="contain", elem_id="post-gallery"), gr.Dataframe( label="Model Predictions", headers=["Arch / Dataset", "By", "AI", "Real", "Label"], datatype=["str", "str", "number", "number", "str"] ), gr.JSON(label="Raw Model Results", visible=False), gr.Markdown(label="Consensus", value="") ], title="Multi-Model Ensemble + Agentic Coordinated Deepfake Detection (Paper in Progress)", description="The detection of AI-generated images has entered a critical inflection point. While existing solutions struggle with outdated datasets and inflated claims, our approach prioritizes agility, community collaboration, and an offensive approach to deepfake detection.", api_name="predict", live=True # Enable streaming ) # def echo_headers(x, request: gr.Request): # print(dict(request.headers)) # return str(dict(request.headers)) def predict(img): """ Predicts whether an image is AI-generated or real using the SOTA Community Forensics model. Args: img (str): Path to the input image file to analyze. Returns: dict: A dictionary containing: - 'Fake Probability' (float): Probability score between 0 and 1 indicating likelihood of being AI-generated - 'Result Description' (str): Human-readable description of the prediction result Example: >>> result = predict("path/to/image.jpg") >>> print(result) {'Fake Probability': 0.002, 'Result Description': 'The image is likely real.'} """ client = Client("aiwithoutborders-xyz/OpenSight-Community-Forensics-Preview") client.view_api() result = client.predict( handle_file(img), api_name="/simple_predict" ) return str(result) community_forensics_preview = gr.Interface( fn=predict, inputs=gr.Image(type="filepath"), outputs=gr.HTML(), # or gr.Markdown() if it's just text title="Quick and simple prediction by our strongest model.", description="No ensemble, no context, no agents, just a quick and simple prediction by our strongest model.", api_name="predict" ) # leaderboard = gr.Interface( # fn=lambda: "# AI Generated / Deepfake Detection Models Leaderboard: Soon™", # inputs=None, # outputs=gr.Markdown(), # title="Leaderboard", # api_name="leaderboard" # ) def simple_prediction(img): """ Quick and simple deepfake or real image prediction by the strongest open-source model on the hub. Args: img (str): The input image to analyze, provided as a file path. Returns: str: The prediction result stringified from dict. Example: `{'Fake Probability': 0.002, 'Result Description': 'The image is likely real.'}` """ client = Client("aiwithoutborders-xyz/OpenSight-Community-Forensics-Preview") client.view_api() client.predict( handle_file(img), api_name="simple_predict" ) simple_predict_interface = gr.Interface( fn=simple_prediction, inputs=gr.Image(type="filepath"), outputs=gr.Text(), title="Quick and simple prediction by our strongest model.", description="No ensemble, no context, no agents, just a quick and simple prediction by our strongest model.", api_name="simple_predict" ) noise_estimation_interface = gr.Interface( fn=noise_estimation, inputs=[gr.Image(type="pil"), gr.Slider(1, 32, value=8, step=1, label="Block Size")], outputs=gr.Image(type="pil"), title="Wavelet-Based Noise Analysis", description="Analyzes image noise patterns using wavelet decomposition. This tool helps detect compression artifacts and artificial noise patterns that may indicate image manipulation. Higher noise levels in specific regions can reveal areas of potential tampering.", api_name="tool_waveletnoise" ) bit_plane_interface = gr.Interface( fn=bit_plane_extractor, inputs=[ gr.Image(type="pil"), gr.Dropdown(["Luminance", "Red", "Green", "Blue", "RGB Norm"], label="Channel", value="Luminance"), gr.Slider(0, 7, value=0, step=1, label="Bit Plane"), gr.Dropdown(["Disabled", "Median", "Gaussian"], label="Filter", value="Disabled") ], outputs=gr.Image(type="pil"), title="Bit Plane Analysis", description="Extracts and visualizes individual bit planes from different color channels. This forensic tool helps identify hidden patterns and artifacts in image data that may indicate manipulation. Different bit planes can reveal inconsistencies in image processing or editing.", api_name="tool_bitplane" ) ela_interface = gr.Interface( fn=ELA, inputs=[ gr.Image(type="pil", label="Input Image"), gr.Slider(1, 100, value=75, step=1, label="JPEG Quality"), gr.Slider(1, 100, value=50, step=1, label="Output Scale (Multiplicative Gain)"), gr.Slider(0, 100, value=20, step=1, label="Output Contrast (Tonality Compression)"), gr.Checkbox(value=False, label="Use Linear Difference"), gr.Checkbox(value=False, label="Grayscale Output") ], outputs=gr.Image(type="pil"), title="Error Level Analysis (ELA)", description="Performs Error Level Analysis to detect re-saved JPEG images, which can indicate tampering. ELA highlights areas of an image that have different compression levels.", api_name="tool_ela" ) gradient_processing_interface = gr.Interface( fn=gradient_processing, inputs=[ gr.Image(type="pil", label="Input Image"), gr.Slider(0, 100, value=90, step=1, label="Intensity"), gr.Dropdown(["Abs", "None", "Flat", "Norm"], label="Blue Mode", value="Abs"), gr.Checkbox(value=False, label="Invert Gradients"), gr.Checkbox(value=False, label="Equalize Histogram") ], outputs=gr.Image(type="pil"), title="Gradient Processing", description="Applies gradient filters to an image to enhance edges and transitions, which can reveal inconsistencies due to manipulation.", api_name="tool_gradient_processing" ) minmax_processing_interface = gr.Interface( fn=minmax_process, inputs=[ gr.Image(type="pil", label="Input Image"), gr.Radio([0, 1, 2, 3, 4], label="Channel (0:Grayscale, 1:Blue, 2:Green, 3:Red, 4:RGB Norm)", value=4), gr.Slider(0, 10, value=2, step=1, label="Radius") ], outputs=gr.Image(type="pil"), title="MinMax Processing", description="Analyzes local pixel value deviations to detect subtle changes in image data, often indicative of digital forgeries.", api_name="tool_minmax_processing" ) # augmentation_tool_interface = gr.Interface( # fn=augment_image, # inputs=[ # gr.Image(label="Upload Image to Augment", sources=['upload', 'webcam'], type='pil'), # gr.CheckboxGroup(["rotate", "add_noise", "sharpen"], label="Augmentation Methods"), # gr.Slider(0, 360, value=0, step=1, label="Rotate Degrees", visible=True), # gr.Slider(0, 100, value=0, step=1, label="Noise Level", visible=True), # gr.Slider(0, 200, value=1, step=1, label="Sharpen Strength", visible=True) # ], # outputs=gr.Image(label="Augmented Image", type='pil'), # title="Image Augmentation Tool", # description="Apply various augmentation techniques to your image.", # api_name="augment_image" # ) # def get_captured_logs(): # # Retrieve all logs from the queue and clear it # logs = list(log_queue) # log_queue.clear() # Clear the queue after retrieving # return "\n".join(logs) demo = gr.TabbedInterface( [ detection_model_eval_playground, community_forensics_preview, noise_estimation_interface, bit_plane_interface, ela_interface, gradient_processing_interface, minmax_processing_interface, # gr.Textbox(label="Agent Logs", interactive=False, lines=5, max_lines=20, autoscroll=True) # New textbox for logs ], [ "Run Ensemble Prediction", "Open-Source SOTA Model", "Wavelet Blocking Noise Estimation", "Bit Plane Values", "Error Level Analysis (ELA)", "Gradient Processing", "MinMax Processing", # "Agent Logs" # New tab title ], title="Deepfake Detection & Forensics Tools", theme=None, ) footerMD = """ ## ⚠️ 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. 3. **UPDATE 6-13-25 (cont.)**: WHILE WE HAVE NOT TAKEN A STANCE ON NSFW AND EXPLICIT CONTENT, PLEASE REFRAIN FROM ... YOUR HUMAN DESIRES UNTIL WE GET THIS PRIVACY SITUATION SORTED OUT. DO NOT BE RECKLESS PLEASE. OUR PAPER WILL BE OUT SOON ON ARXIV WHICH WILL EXPLAIN EVERYTHING WITH DATA-BACKED RESEARCH ON WHY THIS PROJECT IS NEEDED, BUT WE CANNOT DO IT WITHOUT THE HELP OF THE COMMUNITY. 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") with gr.Blocks() as app: demo.render() footer.render() app.queue(max_size=10, default_concurrency_limit=2).launch(mcp_server=True)