File size: 23,698 Bytes
46dd09f | 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 | """
EMOLIPS Pipeline
================
Emotion-Driven Lip-Sync Synthesis Pipeline
Orchestrates:
1. Audio emotion detection (automatic or manual override)
2. Emotion intensity estimation
3. SadTalker talking face generation
4. Emotion-conditioned coefficient modification
5. Output video rendering
Usage:
pipeline = EmolipsPipeline(device="cuda")
pipeline.generate(
audio_path="speech.wav",
image_path="face.jpg",
emotion="happy", # Optional: auto-detected if not specified
intensity=0.7, # Optional: auto-estimated if not specified
output_path="output.mp4"
)
"""
import os
import sys
import subprocess
import shutil
import json
import numpy as np
from pathlib import Path
from typing import Optional, Dict, List
import warnings
warnings.filterwarnings("ignore")
from emotion_module import (
PracticalEmotionModifier,
AudioEmotionDetector,
EmotionIntensityEstimator,
EMOTION_PROFILES
)
class EmolipsPipeline:
"""
Main EMOLIPS inference pipeline.
Wraps SadTalker backbone with emotion conditioning.
"""
def __init__(
self,
sadtalker_dir: str = "./SadTalker",
device: str = "cuda",
checkpoint_dir: str = None
):
self.sadtalker_dir = Path(sadtalker_dir).resolve()
self.device = device
self.checkpoint_dir = checkpoint_dir or str(self.sadtalker_dir / "checkpoints")
# Initialize emotion components
self.emotion_detector = AudioEmotionDetector(device=device)
self.intensity_estimator = EmotionIntensityEstimator()
self.emotion_modifier = PracticalEmotionModifier()
# Verify SadTalker installation
if not self.sadtalker_dir.exists():
print(f"⚠ SadTalker not found at {self.sadtalker_dir}")
print(" Run setup.sh first or specify correct path")
def detect_emotion(self, audio_path: str) -> Dict:
"""Auto-detect emotion from audio."""
print(" [1/4] Detecting emotion from audio...")
result = self.emotion_detector.detect(audio_path)
print(f" Detected: {result['detected_emotion']} "
f"(confidence: {result['confidence']:.2f})")
return result
def estimate_intensity(self, audio_path: str) -> float:
"""Estimate emotion intensity from audio features."""
intensity = self.intensity_estimator.estimate(audio_path)
print(f" Intensity: {intensity:.2f}")
return intensity
def run_sadtalker(
self,
audio_path: str,
image_path: str,
output_dir: str,
expression_scale: float = 1.0,
still_mode: bool = False,
preprocess: str = "crop",
size: int = 256,
pose_style: int = 0
) -> Optional[str]:
"""
Run SadTalker to generate base talking face video.
Returns path to generated video.
"""
print(" [2/4] Running SadTalker backbone...")
# Build SadTalker command
inference_script = self.sadtalker_dir / "inference.py"
cmd = [
sys.executable, str(inference_script),
"--driven_audio", str(audio_path),
"--source_image", str(image_path),
"--result_dir", str(output_dir),
"--expression_scale", str(expression_scale),
"--preprocess", preprocess,
"--size", str(size),
"--pose_style", str(pose_style),
]
if still_mode:
cmd.append("--still")
# Add checkpoint paths
checkpoint_dir = Path(self.checkpoint_dir)
if checkpoint_dir.exists():
cmd.extend(["--checkpoint_dir", str(checkpoint_dir)])
try:
env = os.environ.copy()
env["PYTHONPATH"] = str(self.sadtalker_dir) + ":" + env.get("PYTHONPATH", "")
result = subprocess.run(
cmd,
capture_output=True,
text=True,
cwd=str(self.sadtalker_dir),
env=env,
timeout=300 # 5 min timeout
)
if result.returncode != 0:
print(f" ⚠ SadTalker error: {result.stderr[-500:]}")
return None
# Find generated video
output_path = Path(output_dir)
videos = list(output_path.rglob("*.mp4"))
if videos:
return str(sorted(videos, key=os.path.getmtime)[-1])
return None
except subprocess.TimeoutExpired:
print(" ⚠ SadTalker timed out (>5 min)")
return None
except Exception as e:
print(f" ⚠ SadTalker failed: {e}")
return None
def apply_emotion_postprocess(
self,
video_path: str,
emotion: str,
intensity: float,
output_path: str
) -> str:
"""
Apply emotion-based post-processing to generated video.
This applies subtle facial modifications via:
1. Face landmark detection on each frame
2. Emotion-specific spatial warping
3. Color grading for emotional tone
"""
print(" [3/4] Applying emotion conditioning...")
try:
import cv2
import mediapipe as mp
mp_face_mesh = mp.solutions.face_mesh
face_mesh = mp_face_mesh.FaceMesh(
static_image_mode=False,
max_num_faces=1,
min_detection_confidence=0.5
)
cap = cv2.VideoCapture(video_path)
fps = int(cap.get(cv2.CAP_PROP_FPS))
w = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
h = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
# Temp output (will mux audio later)
temp_path = output_path.replace(".mp4", "_temp.mp4")
fourcc = cv2.VideoWriter_fourcc(*'mp4v')
out = cv2.VideoWriter(temp_path, fourcc, fps, (w, h))
profile = EMOTION_PROFILES.get(emotion, EMOTION_PROFILES["neutral"])
frame_count = 0
total_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
while cap.isOpened():
ret, frame = cap.read()
if not ret:
break
# Apply emotion-specific color grading
frame = self._apply_color_grade(frame, emotion, intensity)
# Apply subtle face warping if emotion is strong
if intensity > 0.3 and emotion != "neutral":
frame = self._apply_face_warp(frame, face_mesh, emotion, intensity)
out.write(frame)
frame_count += 1
cap.release()
out.release()
face_mesh.close()
# Mux original audio back
self._mux_audio(temp_path, video_path, output_path)
# Cleanup temp
if os.path.exists(temp_path):
os.remove(temp_path)
print(f" Processed {frame_count} frames")
return output_path
except ImportError as e:
print(f" ⚠ Post-processing skipped (missing {e}). Copying base video.")
shutil.copy2(video_path, output_path)
return output_path
except Exception as e:
print(f" ⚠ Post-processing error: {e}. Using base video.")
shutil.copy2(video_path, output_path)
return output_path
def _apply_color_grade(
self, frame: np.ndarray, emotion: str, intensity: float
) -> np.ndarray:
"""Apply subtle emotion-specific color grading."""
import cv2
# Very subtle color shifts based on emotion
color_shifts = {
"happy": (5, 5, 15), # Warm (slight yellow)
"sad": (-5, -3, -10), # Cool (slight blue)
"angry": (10, -5, -5), # Warm red
"fear": (-5, -5, 5), # Cool green
"surprise": (5, 5, 5), # Bright
"disgust": (-3, 5, -5), # Sickly green
"neutral": (0, 0, 0),
}
shift = color_shifts.get(emotion, (0, 0, 0))
scale = intensity * 0.5 # Keep it very subtle
adjusted = frame.astype(np.float32)
adjusted[:, :, 0] += shift[0] * scale # B
adjusted[:, :, 1] += shift[1] * scale # G
adjusted[:, :, 2] += shift[2] * scale # R
return np.clip(adjusted, 0, 255).astype(np.uint8)
def _apply_face_warp(
self,
frame: np.ndarray,
face_mesh,
emotion: str,
intensity: float
) -> np.ndarray:
"""
Apply subtle facial warping based on emotion.
Uses MediaPipe landmarks to create emotion-specific deformations.
"""
import cv2
h, w = frame.shape[:2]
rgb = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
results = face_mesh.process(rgb)
if not results.multi_face_landmarks:
return frame
landmarks = results.multi_face_landmarks[0]
# Key landmark indices for warping
# Brow: 70, 63, 105, 66, 107 (left), 336, 296, 334, 293, 300 (right)
# Mouth corners: 61, 291
# Jaw: 152
profile = EMOTION_PROFILES.get(emotion, {})
brow_shift = profile.get("brow_scale", 0) * intensity * 3 # pixels
mouth_shift = profile.get("mouth_scale", 0) * intensity * 2
if abs(brow_shift) < 0.5 and abs(mouth_shift) < 0.5:
return frame # Not enough to notice
# Simple approach: use cv2.remap with subtle displacement
# This is fast and produces decent results
map_x = np.tile(np.arange(w, dtype=np.float32), (h, 1))
map_y = np.tile(np.arange(h, dtype=np.float32).reshape(-1, 1), (1, w))
# Get face center and brow/mouth regions
face_pts = [(int(l.x * w), int(l.y * h)) for l in landmarks.landmark]
# Brow region (top 1/3 of face)
brow_y = face_pts[10][1] # Top of face
nose_y = face_pts[1][1] # Nose tip
brow_region = (brow_y, nose_y)
# Apply brow displacement in brow region
for y_idx in range(max(0, brow_region[0]), min(h, brow_region[1])):
# Gaussian falloff from center of region
region_center = (brow_region[0] + brow_region[1]) // 2
dist = abs(y_idx - region_center) / max(1, (brow_region[1] - brow_region[0]) / 2)
falloff = np.exp(-dist ** 2 * 2)
map_y[y_idx, :] -= brow_shift * falloff
# Apply mouth displacement in lower face
mouth_y = face_pts[13][1] # Upper lip
chin_y = face_pts[152][1] # Chin
mouth_center_x = (face_pts[61][0] + face_pts[291][0]) // 2
for y_idx in range(max(0, mouth_y - 10), min(h, chin_y + 10)):
for x_idx in range(max(0, mouth_center_x - 40), min(w, mouth_center_x + 40)):
dist_y = abs(y_idx - mouth_y) / max(1, (chin_y - mouth_y))
dist_x = abs(x_idx - mouth_center_x) / 40.0
falloff = np.exp(-(dist_y ** 2 + dist_x ** 2) * 2)
map_x[y_idx, x_idx] += mouth_shift * falloff * (1 if x_idx > mouth_center_x else -1)
warped = cv2.remap(frame, map_x, map_y, cv2.INTER_LINEAR)
return warped
def _mux_audio(self, video_path: str, audio_source: str, output_path: str):
"""Combine processed video with original audio."""
try:
subprocess.run([
"ffmpeg", "-y",
"-i", video_path,
"-i", audio_source,
"-c:v", "copy",
"-c:a", "aac",
"-map", "0:v:0",
"-map", "1:a:0",
"-shortest",
output_path
], capture_output=True, timeout=60)
except Exception:
# If ffmpeg fails, just use the video without audio
shutil.copy2(video_path, output_path)
def generate(
self,
audio_path: str,
image_path: str,
emotion: Optional[str] = None,
intensity: Optional[float] = None,
output_path: str = "output.mp4",
expression_scale: float = 1.0,
still_mode: bool = False,
preprocess: str = "crop",
size: int = 256
) -> Dict:
"""
Full EMOLIPS generation pipeline.
Args:
audio_path: Path to speech audio file
image_path: Path to source face image
emotion: Target emotion (auto-detected if None)
intensity: Emotion intensity 0-1 (auto-estimated if None)
output_path: Where to save result
expression_scale: SadTalker expression scale
still_mode: Reduce head motion
preprocess: SadTalker preprocess mode
size: Output resolution
Returns:
Dict with generation metadata
"""
print("=" * 50)
print(" EMOLIPS: Emotion-Driven Lip-Sync Generation")
print("=" * 50)
# Validate inputs
assert os.path.exists(audio_path), f"Audio not found: {audio_path}"
assert os.path.exists(image_path), f"Image not found: {image_path}"
result_meta = {
"audio": audio_path,
"image": image_path,
"output": output_path,
}
# Step 1: Emotion detection
if emotion is None:
detection = self.detect_emotion(audio_path)
emotion = detection["detected_emotion"]
result_meta["emotion_detection"] = detection
else:
print(f" [1/4] Using specified emotion: {emotion}")
result_meta["emotion_detection"] = {"manual": emotion}
# Step 2: Intensity estimation
if intensity is None:
intensity = self.estimate_intensity(audio_path)
else:
print(f" Using specified intensity: {intensity}")
result_meta["emotion"] = emotion
result_meta["intensity"] = intensity
# Adjust SadTalker expression scale based on emotion
emotion_expression_map = {
"neutral": 1.0,
"happy": 1.3,
"sad": 0.9,
"angry": 1.4,
"fear": 1.2,
"surprise": 1.5,
"disgust": 1.1
}
adjusted_scale = expression_scale * emotion_expression_map.get(emotion, 1.0) * (0.5 + 0.5 * intensity)
# Step 3: Run SadTalker
temp_dir = os.path.join(os.path.dirname(output_path) or ".", "temp_sadtalker")
os.makedirs(temp_dir, exist_ok=True)
base_video = self.run_sadtalker(
audio_path=audio_path,
image_path=image_path,
output_dir=temp_dir,
expression_scale=adjusted_scale,
still_mode=still_mode,
preprocess=preprocess,
size=size
)
if base_video is None:
print(" ✗ SadTalker generation failed!")
result_meta["success"] = False
return result_meta
print(f" Base video: {base_video}")
result_meta["base_video"] = base_video
# Step 4: Apply emotion post-processing
final_video = self.apply_emotion_postprocess(
video_path=base_video,
emotion=emotion,
intensity=intensity,
output_path=output_path
)
result_meta["output"] = final_video
result_meta["success"] = True
print(f"\n [4/4] Generation complete!")
print(f" Output: {final_video}")
print(f" Emotion: {emotion} (intensity: {intensity:.2f})")
print("=" * 50)
# Save metadata
meta_path = output_path.replace(".mp4", "_meta.json")
with open(meta_path, "w") as f:
json.dump(result_meta, f, indent=2, default=str)
return result_meta
def generate_all_emotions(
self,
audio_path: str,
image_path: str,
output_dir: str = "outputs",
intensity: float = 0.7,
**kwargs
) -> List[Dict]:
"""
Generate same audio+image across all 7 emotions.
This is the key demo for showing emotion conditioning works.
"""
os.makedirs(output_dir, exist_ok=True)
results = []
emotions = ["neutral", "happy", "sad", "angry", "fear", "surprise", "disgust"]
for emotion in emotions:
print(f"\n{'='*50}")
print(f" Generating: {emotion.upper()}")
print(f"{'='*50}")
out_path = os.path.join(output_dir, f"emolips_{emotion}.mp4")
result = self.generate(
audio_path=audio_path,
image_path=image_path,
emotion=emotion,
intensity=intensity,
output_path=out_path,
**kwargs
)
results.append(result)
# Create comparison grid
self._create_comparison_grid(output_dir, emotions)
return results
def _create_comparison_grid(self, output_dir: str, emotions: List[str]):
"""Create side-by-side comparison video."""
try:
videos = []
for emotion in emotions:
path = os.path.join(output_dir, f"emolips_{emotion}.mp4")
if os.path.exists(path):
videos.append(path)
if len(videos) < 2:
return
# Use ffmpeg to create grid
# 4 videos in a row, 2 rows
filter_parts = []
inputs = []
for i, v in enumerate(videos[:8]): # Max 8
inputs.extend(["-i", v])
filter_parts.append(f"[{i}:v]scale=256:256[v{i}]")
n = len(videos[:8])
cols = min(4, n)
rows = (n + cols - 1) // cols
# Build xstack filter
layout_parts = []
for i in range(min(n, 8)):
x = (i % cols) * 256
y = (i // cols) * 256
layout_parts.append(f"{x}_{y}")
inputs_str = "".join(f"[v{i}]" for i in range(min(n, 8)))
filter_str = ";".join(filter_parts) + f";{inputs_str}xstack=inputs={min(n,8)}:layout={'|'.join(layout_parts)}"
grid_path = os.path.join(output_dir, "comparison_grid.mp4")
subprocess.run(
["ffmpeg", "-y"] + inputs + [
"-filter_complex", filter_str,
"-c:v", "libx264",
"-crf", "23",
grid_path
],
capture_output=True,
timeout=120
)
if os.path.exists(grid_path):
print(f"\n ✓ Comparison grid: {grid_path}")
except Exception as e:
print(f" ⚠ Could not create comparison grid: {e}")
# ============================================================
# STANDALONE MODE (without SadTalker, for testing pipeline)
# ============================================================
class EmolipsStandalone:
"""
Standalone mode that works WITHOUT SadTalker.
Uses MediaPipe face mesh + direct warping for quick demo.
Good for:
- Testing the emotion module independently
- Quick demos without full SadTalker setup
- Verifying the pipeline logic
"""
def __init__(self):
self.emotion_detector = AudioEmotionDetector(device="cpu")
self.intensity_estimator = EmotionIntensityEstimator()
self.emotion_modifier = PracticalEmotionModifier()
def generate_emotion_frames(
self,
image_path: str,
emotion: str,
intensity: float = 0.7,
num_frames: int = 30
) -> List[np.ndarray]:
"""
Generate emotion-modified face frames from a single image.
No audio needed - just shows the emotion transformation.
"""
import cv2
import mediapipe as mp
img = cv2.imread(image_path)
if img is None:
raise ValueError(f"Could not read image: {image_path}")
mp_face_mesh = mp.solutions.face_mesh
face_mesh = mp_face_mesh.FaceMesh(static_image_mode=True, max_num_faces=1)
frames = []
for i in range(num_frames):
# Gradual emotion onset
t = min(1.0, i / (num_frames * 0.3)) # Ramp up in first 30%
current_intensity = intensity * t
frame = img.copy()
# Apply warping
if current_intensity > 0.1:
h, w = frame.shape[:2]
rgb = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
results = face_mesh.process(rgb)
if results.multi_face_landmarks:
profile = EMOTION_PROFILES.get(emotion, {})
brow_shift = profile.get("brow_scale", 0) * current_intensity * 5
mouth_shift = profile.get("mouth_scale", 0) * current_intensity * 4
if abs(brow_shift) > 0.3 or abs(mouth_shift) > 0.3:
map_x = np.tile(np.arange(w, dtype=np.float32), (h, 1))
map_y = np.tile(np.arange(h, dtype=np.float32).reshape(-1, 1), (1, w))
face_pts = [(int(l.x * w), int(l.y * h))
for l in results.multi_face_landmarks[0].landmark]
brow_y = face_pts[10][1]
nose_y = face_pts[1][1]
for y_idx in range(max(0, brow_y), min(h, nose_y)):
center = (brow_y + nose_y) // 2
dist = abs(y_idx - center) / max(1, (nose_y - brow_y) / 2)
falloff = np.exp(-dist ** 2 * 2)
map_y[y_idx, :] -= brow_shift * falloff
frame = cv2.remap(frame, map_x, map_y, cv2.INTER_LINEAR)
# Apply color grading
color_shifts = {
"happy": (5, 5, 15), "sad": (-5, -3, -10),
"angry": (10, -5, -5), "fear": (-5, -5, 5),
"surprise": (5, 5, 5), "disgust": (-3, 5, -5),
"neutral": (0, 0, 0)
}
shift = color_shifts.get(emotion, (0, 0, 0))
adjusted = frame.astype(np.float32)
for c in range(3):
adjusted[:, :, c] += shift[c] * current_intensity * 0.5
frame = np.clip(adjusted, 0, 255).astype(np.uint8)
frames.append(frame)
face_mesh.close()
return frames
def save_demo_video(
self,
image_path: str,
emotions: List[str] = None,
output_dir: str = "outputs",
fps: int = 30,
duration: float = 2.0
):
"""Save emotion demo videos from a single face image."""
import cv2
os.makedirs(output_dir, exist_ok=True)
if emotions is None:
emotions = ["neutral", "happy", "sad", "angry", "fear", "surprise", "disgust"]
num_frames = int(fps * duration)
for emotion in emotions:
print(f" Generating {emotion}...")
frames = self.generate_emotion_frames(image_path, emotion, 0.7, num_frames)
out_path = os.path.join(output_dir, f"demo_{emotion}.mp4")
h, w = frames[0].shape[:2]
out = cv2.VideoWriter(out_path, cv2.VideoWriter_fourcc(*'mp4v'), fps, (w, h))
for f in frames:
out.write(f)
out.release()
print(f" ✓ {out_path}")
if __name__ == "__main__":
print("EMOLIPS Pipeline module loaded.")
print("Use EmolipsPipeline for full generation or EmolipsStandalone for quick demo.")
|