Implement ear preservation for Poisson blending.

- Added functionality to preserve target's ears when using Poisson blending.
- Introduced new global variables in `globals.py`:
  - `preserve_target_ears` (boolean flag)
  - `ear_width_ratio`
  - `ear_height_ratio`
  - `ear_vertical_offset_ratio`
  - `ear_horizontal_overlap_ratio`
- Added command-line argument `--preserve-ears` to `core.py`.
- Modified `face_swapper.py` to adjust the Poisson blending mask by
  subtracting calculated ear regions if `preserve_target_ears` is true.
  This aims to reduce artifacts around ears by using the original target
  frame's ear content.

modules/core.py

@@ -44,6 +44,7 @@ def parse_args() -> None:
     program.add_argument('--map-faces', help='map source target faces', dest='map_faces', action='store_true', default=False)
     program.add_argument('--mouth-mask', help='mask the mouth region', dest='mouth_mask', action='store_true', default=False)
     program.add_argument('--poisson-blending', help='use Poisson blending for smoother face integration', dest='poisson_blending', action='store_true', default=False)
+    program.add_argument('--preserve-ears', help='attempt to preserve target ears by modifying the blend mask', dest='preserve_ears', action='store_true', default=False)
     program.add_argument('--video-encoder', help='adjust output video encoder', dest='video_encoder', default='libx264', choices=['libx264', 'libx265', 'libvpx-vp9'])
     program.add_argument('--video-quality', help='adjust output video quality', dest='video_quality', type=int, default=18, choices=range(52), metavar='[0-51]')
     program.add_argument('-l', '--lang', help='Ui language', default="en")
@@ -74,6 +75,7 @@ def parse_args() -> None:
     modules.globals.color_correction = args.color_correction
     modules.globals.mouth_mask = args.mouth_mask
     modules.globals.use_poisson_blending = args.poisson_blending
+    modules.globals.preserve_target_ears = args.preserve_ears
     modules.globals.nsfw_filter = args.nsfw_filter
     modules.globals.map_faces = args.map_faces
     modules.globals.video_encoder = args.video_encoder

modules/globals.py

@@ -43,3 +43,8 @@ mask_down_size = 0.50
 mask_size = 1
 use_poisson_blending = False # Added for Poisson blending
 poisson_blending_feather_amount = 5 # Feathering for the mask before Poisson blending
+preserve_target_ears = False # Flag to enable preserving target's ears
+ear_width_ratio = 0.18 # Width of the ear exclusion box as a ratio of face bbox width
+ear_height_ratio = 0.35 # Height of the ear exclusion box as a ratio of face bbox height
+ear_vertical_offset_ratio = 0.20 # Vertical offset of the ear box from top of face bbox
+ear_horizontal_overlap_ratio = 0.03 # How much the ear exclusion zone can overlap into the face bbox

modules/processors/frame/face_swapper.py

@@ -172,12 +172,41 @@ def swap_face(source_face: Face, target_face: Face, temp_frame: Frame) -> Frame:
             else:
                 logging.error("No bbox or landmarks available for Poisson mask. Blending will be skipped.")
 
+        # Subtract ear regions if preserve_target_ears is enabled
+        if modules.globals.preserve_target_ears and np.any(face_mask_for_blending > 0):
+            mfx1, mfy1, mfx2, mfy2 = target_face.bbox.astype(int)
+            mfw = mfx2 - mfx1
+            mfh = mfy2 - mfy1
+
+            ear_w = int(mfw * modules.globals.ear_width_ratio)
+            ear_h = int(mfh * modules.globals.ear_height_ratio)
+            ear_v_offset = int(mfh * modules.globals.ear_vertical_offset_ratio)
+            ear_overlap = int(mfw * modules.globals.ear_horizontal_overlap_ratio)
+
+            # Person's Right Ear (image left side of face bbox)
+            # This region in face_mask_for_blending will be set to 0
+            rex1 = max(0, mfx1 - ear_w + ear_overlap)
+            rey1 = max(0, mfy1 + ear_v_offset)
+            rex2 = min(temp_frame.shape[1], mfx1 + ear_overlap) # Extends slightly into face bbox for smoother transition
+            rey2 = min(temp_frame.shape[0], rey1 + ear_h)
+            if rex1 < rex2 and rey1 < rey2:
+                cv2.rectangle(face_mask_for_blending, (rex1, rey1), (rex2, rey2), 0, -1)
+
+            # Person's Left Ear (image right side of face bbox)
+            lex1 = max(0, mfx2 - ear_overlap)
+            ley1 = max(0, mfy1 + ear_v_offset)
+            lex2 = min(temp_frame.shape[1], mfx2 + ear_w - ear_overlap)
+            ley2 = min(temp_frame.shape[0], ley1 + ear_h)
+            if lex1 < lex2 and ley1 < ley2:
+                cv2.rectangle(face_mask_for_blending, (lex1, ley1), (lex2, ley2), 0, -1)
+
         # Feather the mask to smooth edges for Poisson blending
-        feather_amount = modules.globals.poisson_blending_feather_amount
-        if feather_amount > 0:
-            # Ensure kernel size is odd
-            kernel_size = 2 * feather_amount + 1
-            face_mask_for_blending = cv2.GaussianBlur(face_mask_for_blending, (kernel_size, kernel_size), 0)
+        if np.any(face_mask_for_blending > 0): # Only feather if there's a mask
+            feather_amount = modules.globals.poisson_blending_feather_amount
+            if feather_amount > 0:
+                # Ensure kernel size is odd
+                kernel_size = 2 * feather_amount + 1
+                face_mask_for_blending = cv2.GaussianBlur(face_mask_for_blending, (kernel_size, kernel_size), 0)
 
         # Calculate the center of the target face bbox for seamlessClone
         if hasattr(target_face, 'bbox'):