File size: 21,569 Bytes
2d1cb2c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import argparse
import json
import re
from pathlib import Path
from typing import List, Dict, Tuple, Optional, Any
import math

import numpy as np
import torch
import torch.nn as nn
from tqdm import tqdm
from PIL import Image, ImageOps
import matplotlib.cm as cm
import torchvision.transforms as T
from torchvision.transforms.functional import InterpolationMode

# Module 1 Imports
from transformers import AutoImageProcessor, AutoModelForImageClassification

# Module 2 Imports (InternVL)
from transformers import AutoModel, AutoTokenizer

# -----------------------------------------------------------------------------
# Configuration & Constants
# -----------------------------------------------------------------------------
IMG_EXTS = {".jpg", ".jpeg", ".png", ".webp", ".bmp", ".tif", ".tiff"}

# -----------------------------------------------------------------------------
# DYNAMIC PROMPT TEMPLATE
# -----------------------------------------------------------------------------
VLM_SYSTEM_PROMPT_TEMPLATE = """
Role: You are a Digital Forensics Expert.

Input Context:
Image-1: The suspect image.
Image-2: A Grad-CAM heatmap (Red = Pixel Artifacts detected).
Forensic Score: {authenticity_score:.2f} (0.0=Clear, 1.0=Flagged).

Technical Status: {status_msg}

Your Mission: {mission_msg}

Step-by-Step Analysis:
1. Physics Check: Do shadows, reflections, and lighting match the environment?
2. Biological Integrity: Check for wax-like skin, asymmetrical eyes, or blending lines on the neck.
3. Logic Check: Are there impossible geometries or structural errors?

Output Requirements:
Output ONLY a JSON object.
"manipulation_type": Select the best fit from: {allowed_options}
"vlm_reasoning": {reasoning_instruction}

Constraint: {constraint_msg}
"""

# -----------------------------------------------------------------------------
# Utils
# -----------------------------------------------------------------------------
def list_images(folder: Path) -> List[Path]:
    return sorted([p for p in folder.rglob("*") if p.is_file() and p.suffix.lower() in IMG_EXTS])

def load_rgb(path: Path) -> Image.Image:
    img = Image.open(path)
    img = ImageOps.exif_transpose(img)
    if img.mode != "RGB":
        img = img.convert("RGB")
    return img

def resize_pad_square(img: Image.Image, size: int) -> Image.Image:
    w, h = img.size
    if w <= 0 or h <= 0:
        return img.resize((size, size), resample=Image.BICUBIC)
    scale = size / float(max(w, h))
    new_w = max(1, int(round(w * scale)))
    new_h = max(1, int(round(h * scale)))
    img = img.resize((new_w, new_h), resample=Image.BICUBIC)
    pad_left = (size - new_w) // 2
    pad_top = (size - new_h) // 2
    pad_right = size - new_w - pad_left
    pad_bottom = size - new_h - pad_top
    img = ImageOps.expand(img, border=(pad_left, pad_top, pad_right, pad_bottom), fill=0)
    return img

# -----------------------------------------------------------------------------
# Module 1: Forensic Detector Helpers
# -----------------------------------------------------------------------------
def get_norm_from_processor(processor) -> Tuple[List[float], List[float], float]:
    mean = getattr(processor, "image_mean", [0.485, 0.456, 0.406])
    std = getattr(processor, "image_std", [0.229, 0.224, 0.225])
    rescale_factor = getattr(processor, "rescale_factor", 1.0 / 255.0)
    return list(mean), list(std), float(rescale_factor)

def preprocess_one(img: Image.Image, size: int, mean: List[float], std: List[float], rescale_factor: float) -> Tuple[torch.Tensor, Image.Image]:
    img_sq = resize_pad_square(img, size)
    arr = np.array(img_sq).astype(np.float32)
    arr = arr * rescale_factor
    arr = np.transpose(arr, (2, 0, 1))
    x = torch.from_numpy(arr)
    m = torch.tensor(mean, dtype=torch.float32)[:, None, None]
    s = torch.tensor(std, dtype=torch.float32)[:, None, None]
    x = (x - m) / s
    return x, img_sq

def preprocess_batch(imgs: List[Image.Image], size: int, mean: List[float], std: List[float], rescale_factor: float) -> torch.Tensor:
    xs = []
    for im in imgs:
        x, _ = preprocess_one(im, size, mean, std, rescale_factor)
        xs.append(x)
    return torch.stack(xs, dim=0)

@torch.inference_mode()
def forward_fake_prob(model, pixel_values: torch.Tensor, fake_idx: int) -> torch.Tensor:
    out = model(pixel_values=pixel_values)
    logits = out.logits
    if logits.shape[-1] == 1:
        prob = torch.sigmoid(logits[:, 0])
    else:
        prob = torch.softmax(logits, dim=-1)[:, fake_idx]
    return prob

@torch.inference_mode()
def predict_probs_batch(model, paths: List[Path], device: torch.device, size: int, mean: List[float], std: List[float], rescale_factor: float, fake_idx: int, use_tta: bool) -> List[float]:
    raw_images = [load_rgb(p) for p in paths]
    if not use_tta:
        pv = preprocess_batch(raw_images, size, mean, std, rescale_factor).to(device)
        probs = forward_fake_prob(model, pv, fake_idx)
        return probs.detach().cpu().tolist()
    
    # Base
    pv_base = preprocess_batch(raw_images, size, mean, std, rescale_factor).to(device)
    probs_sum = forward_fake_prob(model, pv_base, fake_idx)

    # 4 Quadrants
    imgs_tl, imgs_tr, imgs_bl, imgs_br = [], [], [], []
    for img in raw_images:
        w, h = img.size
        mid_w, mid_h = w // 2, h // 2
        imgs_tl.append(img.crop((0, 0, mid_w, mid_h)))
        imgs_tr.append(img.crop((mid_w, 0, w, mid_h)))
        imgs_bl.append(img.crop((0, mid_h, mid_w, h)))
        imgs_br.append(img.crop((mid_w, mid_h, w, h)))

    for quad_imgs in (imgs_tl, imgs_tr, imgs_bl, imgs_br):
        pv_q = preprocess_batch(quad_imgs, size, mean, std, rescale_factor).to(device)
        probs_sum = probs_sum + forward_fake_prob(model, pv_q, fake_idx)

    probs = probs_sum / 5.0
    return probs.detach().cpu().tolist()

# -----------------------------------------------------------------------------
# Grad-CAM
# -----------------------------------------------------------------------------
class GradCAM:
    def __init__(self, model: nn.Module, target_layer: nn.Module):
        self.model = model
        self.target_layer = target_layer
        self.activations = None
        self.gradients = None
        self._fwd = target_layer.register_forward_hook(self._forward_hook)
        self._bwd = target_layer.register_full_backward_hook(self._backward_hook)

    def close(self):
        self._fwd.remove()
        self._bwd.remove()

    def _forward_hook(self, module, inp, out):
        self.activations = out

    def _backward_hook(self, module, grad_input, grad_output):
        self.gradients = grad_output[0]

    def __call__(self, pixel_values: torch.Tensor, class_index: int) -> torch.Tensor:
        self.model.zero_grad(set_to_none=True)
        out = self.model(pixel_values=pixel_values)
        logits = out.logits
        if logits.shape[-1] == 1:
            score = logits[:, 0]
        else:
            score = logits[:, class_index]
        score.sum().backward(retain_graph=False)
        
        acts = self.activations
        grads = self.gradients
        weights = grads.mean(dim=(2, 3), keepdim=True)
        cam = (weights * acts).sum(dim=1)
        cam = torch.relu(cam)
        cam_min = cam.amin(dim=(1, 2), keepdim=True)
        cam_max = cam.amax(dim=(1, 2), keepdim=True)
        cam = (cam - cam_min) / (cam_max - cam_min + 1e-6)
        return cam[0].detach()

def make_overlay(pil_img: Image.Image, cam_01: np.ndarray, alpha: float = 0.45) -> Image.Image:
    cam_01 = np.clip(cam_01, 0.0, 1.0)
    heat = cm.get_cmap("jet")(cam_01)[:, :, :3]
    heat_u8 = (heat * 255.0).astype(np.uint8)
    base = np.array(pil_img).astype(np.uint8)
    if heat_u8.shape[:2] != base.shape[:2]:
        heat_pil = Image.fromarray(heat_u8).resize((base.shape[1], base.shape[0]), Image.BILINEAR)
        heat_u8 = np.array(heat_pil)
    
    overlay = (base * (1.0 - alpha) + heat_u8 * alpha).astype(np.uint8)
    return Image.fromarray(overlay)

# -----------------------------------------------------------------------------
# Module 2: InternVL Preprocessing Utilities
# -----------------------------------------------------------------------------
def build_transform(input_size):
    MEAN, STD = (0.485, 0.456, 0.406), (0.229, 0.224, 0.225)
    transform = T.Compose([
        T.Lambda(lambda img: img.convert('RGB') if img.mode != 'RGB' else img),
        T.Resize((input_size, input_size), interpolation=InterpolationMode.BICUBIC),
        T.ToTensor(),
        T.Normalize(mean=MEAN, std=STD)
    ])
    return transform

def find_closest_aspect_ratio(aspect_ratio, target_ratios, width, height, image_size):
    best_ratio_diff = float('inf')
    best_ratio = (1, 1)
    area = width * height
    for ratio in target_ratios:
        target_aspect_ratio = ratio[0] / ratio[1]
        ratio_diff = abs(aspect_ratio - target_aspect_ratio)
        if ratio_diff < best_ratio_diff:
            best_ratio_diff = ratio_diff
            best_ratio = ratio
        elif ratio_diff == best_ratio_diff:
            if area > 0.5 * image_size * image_size * ratio[0] * ratio[1]:
                best_ratio = ratio
    return best_ratio

def dynamic_preprocess(image, min_num=1, max_num=6, image_size=448, use_thumbnail=True):
    orig_width, orig_height = image.size
    aspect_ratio = orig_width / orig_height
    target_ratios = set(
        (i, j) for n in range(min_num, max_num + 1) for i in range(1, n + 1) for j in range(1, n + 1) if
        i * j <= max_num and i * j >= min_num)
    target_ratios = sorted(list(target_ratios), key=lambda x: x[0] * x[1])
    target_aspect_ratio = find_closest_aspect_ratio(
        aspect_ratio, target_ratios, orig_width, orig_height, image_size)
    target_width = image_size * target_aspect_ratio[0]
    target_height = image_size * target_aspect_ratio[1]
    blocks = target_aspect_ratio[0] * target_aspect_ratio[1]
    resized_img = image.resize((target_width, target_height))
    processed_images = []
    for i in range(blocks):
        box = (
            (i % (target_width // image_size)) * image_size,
            (i // (target_width // image_size)) * image_size,
            ((i % (target_width // image_size)) + 1) * image_size,
            ((i // (target_width // image_size)) + 1) * image_size
        )
        split_img = resized_img.crop(box)
        processed_images.append(split_img)
    if use_thumbnail and len(processed_images) > 1:
        thumbnail_img = image.resize((image_size, image_size))
        processed_images.append(thumbnail_img)
    return processed_images

# -----------------------------------------------------------------------------
# Module 2: VLM Logic (InternVL)
# -----------------------------------------------------------------------------
def load_internvl(model_name: str, cache_dir: str):
    print(f"Loading VLM: {model_name}...")
    tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True, cache_dir=cache_dir)
    # Using float16 or bfloat16 for efficiency
    dtype = torch.bfloat16 if torch.cuda.is_bf16_supported() else torch.float16
    
    # === MULTI-GPU CHANGE ===
    # Using device_map="auto" lets Hugging Face Accelerate split layers across GPUs 0,1,2,3
    print("Dispatching model across available GPUs (device_map='auto')...")
    model = AutoModel.from_pretrained(
        model_name, 
        trust_remote_code=True, 
        torch_dtype=dtype,
        low_cpu_mem_usage=True,
        cache_dir=cache_dir,
        use_flash_attn=False,
        device_map="auto"  # This enables Multi-GPU usage
    ).eval()
    
    return tokenizer, model

def run_vlm_audit(tokenizer, model, orig_path: str, cam_path: str, score: float) -> Dict[str, Any]:
    """
    Constructs the prompt and runs inference on InternVL.
    """
    
    # === COMPLEMENTARY AUDIT STRATEGY ===
    if score >= 0.5:
        # High Likelihood of Fake (Pixel Detector found artifacts)
        status_msg = "Pixel-level artifacts DETECTED. The image is likely manipulated."
        mission_msg = "Confirm the specific type of manipulation. Does the red heatmap align with semantic errors?"
        allowed_options = "['In-painting', 'Full Synthesis', 'Face Swap', 'Filter']"
        reasoning_instruction = "Explain which specific feature (eyes, neck, shadow) aligns with the heatmap to prove the manipulation."
        constraint_msg = "You MUST classify the type of manipulation. Do not choose 'None' unless the pixel detector is clearly hallucinating (extremely rare)."
    
    else:
        # Low Likelihood (Pixel Detector is happy)
        status_msg = "Pixel-level artifacts NOT detected. The image passed the noise/frequency check."
        mission_msg = "Hunt for 'Semantic Impossibilities' that the pixel detector missed (e.g., bad physics, lighting errors). If the physics and logic are perfect, mark as None."
        allowed_options = "['None', 'In-painting', 'Full Synthesis', 'Face Swap', 'Filter']"
        reasoning_instruction = "If authentic, state 'No semantic anomalies found'. If fake, explain the physical impossibility (e.g. 'shadows go wrong direction') that proves it despite clean pixels."
        constraint_msg = "Prefer 'None' if the image looks natural. Only flag if you find a logical or physical contradiction."

    # Fill template
    prompt_text = VLM_SYSTEM_PROMPT_TEMPLATE.format(
        authenticity_score=score,
        status_msg=status_msg,
        mission_msg=mission_msg,
        allowed_options=allowed_options,
        reasoning_instruction=reasoning_instruction,
        constraint_msg=constraint_msg
    )
    
    # 2. Load and Process Images
    img1 = load_rgb(Path(orig_path))
    img2 = load_rgb(Path(cam_path))
    
    transform = build_transform(input_size=448)
    
    # Process both images into tiles
    tiles1 = dynamic_preprocess(img1, image_size=448, use_thumbnail=True, max_num=6)
    tiles2 = dynamic_preprocess(img2, image_size=448, use_thumbnail=True, max_num=6)
    
    # Stack pixels
    # Note: We must move tensors to model.device (which is usually the device of the first layer)
    pixel_values1 = [transform(t) for t in tiles1]
    pixel_values2 = [transform(t) for t in tiles2]
    
    # Move to GPU
    target_device = model.device
    pixel_values = torch.stack(pixel_values1 + pixel_values2).to(torch.bfloat16 if torch.cuda.is_bf16_supported() else torch.float16).to(target_device)
    
    # 3. Construct Question
    question = f"Image-1: <image>\nImage-2: <image>\n{prompt_text}"
    
    generation_config = dict(max_new_tokens=512, do_sample=False)

    try:
        response = model.chat(tokenizer, pixel_values, question, generation_config)
    except Exception as e:
        return {"manipulation_type": "Error", "vlm_reasoning": f"VLM Inference Error: {e}"}

    # 4. Extract JSON
    try:
        json_match = re.search(r"\{.*\}", response, re.DOTALL)
        if json_match:
            json_str = json_match.group(0)
            data = json.loads(json_str)
            return data
        else:
            return {"manipulation_type": "Unknown", "vlm_reasoning": response}
    except Exception as e:
        return {"manipulation_type": "Error", "vlm_reasoning": f"Failed to parse JSON: {response}"}

# -----------------------------------------------------------------------------
# Main Pipeline
# -----------------------------------------------------------------------------
def main():
    ap = argparse.ArgumentParser()
    ap.add_argument("--input_dir", type=str, required=True)
    ap.add_argument("--output_file", type=str, default="predictions.json")
    ap.add_argument("--model_id", type=str, default="buildborderless/CommunityForensics-DeepfakeDet-ViT")
    ap.add_argument("--vlm_id", type=str, default="OpenGVLab/InternVL3_5-30B-A3B-MPO")
    ap.add_argument("--cache_dir", type=str, default="./")
    ap.add_argument("--device", type=str, default="auto")
    ap.add_argument("--batch_size", type=int, default=8)
    ap.add_argument("--tta", action="store_true", help="Enable TTA for ViT")
    args = ap.parse_args()

    # Device Setup (for ViT)
    # InternVL handles its own device map, but ViT needs explicit device
    if args.device == "auto":
        # Put ViT on the first GPU explicitly
        device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
    else:
        device = torch.device(args.device)
    print(f"Using device for Module 1 (ViT): {device}")

    input_dir = Path(args.input_dir)
    out_file = Path(args.output_file)
    cam_dir = out_file.parent / "gradcam"
    cam_dir.mkdir(parents=True, exist_ok=True)

    # ---------------------------
    # 1. Load Module 1 (ViT)
    # ---------------------------
    print(f"--- Loading Module 1: {args.model_id} ---")
    processor = AutoImageProcessor.from_pretrained(args.model_id, cache_dir=args.cache_dir)
    vit_model = AutoModelForImageClassification.from_pretrained(args.model_id, cache_dir=args.cache_dir).to(device).eval()
    
    mean, std, rescale_factor = get_norm_from_processor(processor)
    size = 384 
    try:
        size = vit_model.config.image_size
        if isinstance(size, (tuple, list)): size = size[0]
    except:
        pass
    
    fake_idx = 1
    if hasattr(vit_model.config, "label2id"):
        for k, v in vit_model.config.label2id.items():
            if "fake" in k.lower(): fake_idx = v; break

    # Setup GradCAM
    target_layer = None
    for name, module in vit_model.named_modules():
        if "patch_embeddings.projection" in name and isinstance(module, nn.Conv2d):
            target_layer = module
            break
    if target_layer is None: 
        for module in vit_model.modules():
            if isinstance(module, nn.Conv2d): target_layer = module
    
    gradcam = GradCAM(vit_model, target_layer) if target_layer else None
    print(f"GradCAM Layer: {target_layer}")

    # ---------------------------
    # 2. Run Module 1 Inference
    # ---------------------------
    paths = list_images(input_dir)
    print(f"Found {len(paths)} images. Running Forensic Scan...")
    
    results_map = {} 

    for i in tqdm(range(0, len(paths), args.batch_size), desc="ViT Scanning"):
        batch_paths = paths[i:i+args.batch_size]
        scores = predict_probs_batch(
            model=vit_model,
            paths=batch_paths,
            device=device,
            size=size,
            mean=mean,
            std=std,
            rescale_factor=rescale_factor,
            fake_idx=fake_idx,
            use_tta=args.tta
        )
        for p, s in zip(batch_paths, scores):
            results_map[p] = {"score": s, "cam_path": None}

    print("Generating Heatmaps for ALL images...")
    for p, data in tqdm(results_map.items(), desc="Grad-CAM Gen"):
        if gradcam:
            img = load_rgb(p)
            x, img_sq = preprocess_one(img, size, mean, std, rescale_factor)
            pv = x.unsqueeze(0).to(device)
            pv.requires_grad_(True)
            
            try:
                cam = gradcam(pv, class_index=fake_idx)
                cam_np = cam.cpu().numpy()
                W, H = img_sq.size
                cam_pil = Image.fromarray((cam_np * 255).astype(np.uint8)).resize((W, H), Image.BILINEAR)
                cam_norm = np.array(cam_pil) / 255.0
                
                overlay = make_overlay(img_sq, cam_norm)
                
                rel_name = p.relative_to(input_dir)
                save_path = cam_dir / (str(rel_name).replace("/", "_") + ".png")
                save_path.parent.mkdir(parents=True, exist_ok=True)
                overlay.save(save_path)
                
                data["cam_path"] = str(save_path.absolute())
            except Exception as e:
                print(f"CAM Error on {p}: {e}")

    if gradcam: gradcam.close()
    
    # === CRITICAL MEMORY CLEANUP ===
    del vit_model, gradcam, processor
    torch.cuda.empty_cache()
    # ===============================

    # ---------------------------
    # 3. Load Module 2 (InternVL)
    # ---------------------------
    print(f"--- Loading Module 2: {args.vlm_id} ---")
    # Pass only the cache_dir, device is handled auto
    tokenizer, vlm_model = load_internvl(args.vlm_id, args.cache_dir)

    # ---------------------------
    # 4. Fusion & Audit
    # ---------------------------
    final_json = []
    
    print("Running VLM Semantic Audit on ALL images...")
    for p, data in tqdm(results_map.items(), desc="VLM Reasoning"):
        score = data["score"]
        cam_path = data["cam_path"]
        
        rel_name = str(p.relative_to(input_dir))
        
        # Default Fallbacks
        m_type = "None"
        reasoning = "Forensic score is low and no anomalies detected."
        
        if cam_path:
            vlm_out = run_vlm_audit(
                tokenizer, 
                vlm_model, 
                orig_path=str(p.absolute()), 
                cam_path=cam_path, 
                score=score
            )
            m_type = vlm_out.get("manipulation_type", "Unknown")
            reasoning = vlm_out.get("vlm_reasoning", "VLM failed to reason.")
        else:
            reasoning = "VLM Skipped (Missing Heatmap)"

        final_json.append({
            "image_name": rel_name,
            "authenticity_score": float(score),
            "manipulation_type": m_type,
            "vlm_reasoning": reasoning
        })

    with open(out_file, "w") as f:
        json.dump(final_json, f, indent=2)
    
    print(f"Done! Predictions saved to {out_file}")

if __name__ == "__main__":
    main()