File size: 33,505 Bytes
63aad13 acf0228 63aad13 1146644 29f83d8 63aad13 bc355a9 5342779 bc355a9 63aad13 5342779 63aad13 bc355a9 63aad13 e1eac06 0c436d0 e1eac06 63aad13 bc355a9 63aad13 be96dd0 63aad13 0c436d0 63aad13 e1eac06 ed7026a e1eac06 63aad13 be96dd0 63aad13 088ff8c 63aad13 be96dd0 63aad13 be96dd0 63aad13 be96dd0 63aad13 be96dd0 63aad13 be96dd0 63aad13 acf0228 63aad13 bc355a9 63aad13 2be263d 63aad13 c3c2134 5342779 30f141e 63aad13 5342779 63aad13 1146644 63aad13 c3c2134 63aad13 c56a0f7 63aad13 c56a0f7 63aad13 c56a0f7 63aad13 c56a0f7 63aad13 c56a0f7 63aad13 ed7026a 63aad13 ed7026a 63aad13 ed7026a 63aad13 5c2bf26 63aad13 c56a0f7 63aad13 5342779 63aad13 30f141e 63aad13 be96dd0 63aad13 8159928 5342779 c56a0f7 63aad13 e1eac06 63aad13 e1eac06 737fdca e1eac06 ed7026a 63aad13 e1eac06 acf0228 63aad13 5342779 e1eac06 63aad13 acf0228 e1eac06 63aad13 e1eac06 63aad13 5342779 63aad13 e1eac06 acf0228 29f83d8 acf0228 63aad13 62180cb 5342779 63aad13 acf0228 7cd1bf7 63aad13 7cd1bf7 63aad13 acf0228 7cd1bf7 a677ac6 48e0567 63aad13 a77a8d3 63aad13 e01838e e1eac06 0c436d0 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 |
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
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
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": "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"
# )
# --- ONNX Quantized Model Example ---
ONNX_QUANTIZED_MODEL_PATH = "./models/model_1_quantized.onnx"
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')
transform = transforms.Compose([
transforms.Resize((256, 256)),
transforms.ToTensor(),
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(preprocessed_image_np):
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 = 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]
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: {e}")
# Return a structure consistent with other model errors
return {"logits": np.array([]), "probabilities": np.array([])}
def postprocess_onnx_output(onnx_output, class_names):
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))}
else:
logger.warning("ONNX post-processing failed or class names mismatch.")
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"
)
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]
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>"
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
)
# 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("", 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) |