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 # Assuming these are available from your utils and agents directories # You might need to adjust paths or copy these functions/classes if they are not directly importable. 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 import torch import json from huggingface_hub import CommitScheduler from dotenv import load_dotenv # Configure logging logging.basicConfig(level=logging.INFO) logger = logging.getLogger(__name__) os.environ['HF_HUB_CACHE'] = './models' 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": "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": "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": ['AI', 'Real'], "model_4": ['AI', 'Real'], "model_5": ['Realism', 'Deepfake'], "model_6": ['ai_gen', 'human'], "model_7": ['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']: 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 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" ) 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. 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) 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, "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 = [] # 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() monitor_agent.monitor_prediction( model_id, result["Label"], 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(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(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, None, 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, } 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, bitplane_image] 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=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"]: 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 ) community_forensics_preview = gr.Interface( fn=lambda: gr.load("aiwithoutborders-xyz/OpenSight-Community-Forensics-Preview", src="spaces"), 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="quick_predict" ) # leaderboard = gr.Interface( # fn=lambda: "# AI Generated / Deepfake Detection Models Leaderboard: Soon™", # inputs=None, # outputs=gr.Markdown(), # title="Leaderboard", # api_name="leaderboard" # ) # 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" ) def augment_image_interface(img, augment_methods, rotate_degrees, noise_level, sharpen_strength): if img is None: raise gr.Error("No image provided for augmentation. Please upload an image.") # Ensure image is PIL Image and in RGB format if not isinstance(img, Image.Image): try: img = Image.fromarray(img) except Exception as e: raise gr.Error(f"Could not convert input to PIL Image: {e}") if img.mode != 'RGB': img = img.convert('RGB') augmented_img, _ = augment_image(img, augment_methods, rotate_degrees, noise_level, sharpen_strength) return augmented_img augmentation_tool_interface = gr.Interface( fn=augment_image_interface, 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" ) demo = gr.TabbedInterface( [ detection_model_eval_playground, community_forensics_preview, noise_estimation_interface, bit_plane_interface, ela_interface, gradient_processing_interface, minmax_processing_interface, augmentation_tool_interface ], [ "Run Ensemble Prediction", "Community Model", "Wavelet Blocking Noise Estimation", "Bit Plane Values", "Error Level Analysis (ELA)", "Gradient Processing", "MinMax Processing", "Image Augmentation" ], title="Deepfake Detection & Forensics Tools", theme=None, ) if __name__ == "__main__": demo.launch(mcp_server=True)