v1.9.117: Seamless Transition & Lite Bias Recovery

aiia_ditto_nodes.py

@@ -754,7 +754,7 @@ class AIIA_DittoSampler:
             # Only applied when alpha < 1.0.
             
             idle_amp = 4.5 
-            DECAY_FRAMES = 25 # ~1.0s
+            DECAY_FRAMES = 50 # [v1.9.117] Sync with Core warping (80 frames) for seamless handover
             
             for i in range(num_frames):
                 alpha = float(dataset_alpha[i])

libs/Ditto/core/atomic_components/audio2motion.py

@@ -259,7 +259,7 @@ class Audio2Motion:
              
              # Apply decaying warp to the new sequence
              # We want to return to the model's path over ~50 frames (v1.9.107: longer decay)
-             decay_len = 50
+             decay_len = 80
              warp_len = min(decay_len, pred_kp_seq.shape[1])
              weights = np.linspace(1.0, 0.0, warp_len).reshape(1, -1, 1)
              

libs/Ditto/core/atomic_components/motion_stitch.py

@@ -681,14 +681,19 @@ class MotionStitch:
             exp_reshaped[:, _squint_fix] = src_reshaped[:, _squint_fix]
             x_d_info["exp"] = exp_reshaped.reshape(1, -1)
 
-        # [v1.9.116] DIAGNOSTIC: Temporarily disabled mouth bias to trace facial mutation
-        # bias_alpha = kwargs.get("vad_alpha", 0.0)
-        # lower_lip = [17, 19, 20]
-        # corners = [7, 8]
-        # exp_reshaped = x_d_info["exp"].reshape(-1, 21, 3)
-        # exp_reshaped[:, lower_lip, 1] += 0.05 * bias_alpha
-        # exp_reshaped[:, corners, 1] += 0.015 * bias_alpha 
-        # x_d_info["exp"] = exp_reshaped.reshape(1, -1)
+        # [v1.9.117] Lite Mouth Opening Bias (Applied BEFORE EMA for temporal smoothing)
+        # We use a 60% strength (+0.03) compared to v1.9.112 to avoid facial mutation.
+        bias_alpha = kwargs.get("vad_alpha", 0.0)
+        lower_lip = [17, 19, 20]
+        corners = [7, 8]
+        
+        # Point to spatial indexing: exp is (1, 21, 3)
+        exp_reshaped = x_d_info["exp"].reshape(-1, 21, 3)
+        # Center lower lip gets lite bias (+0.03)
+        exp_reshaped[:, lower_lip, 1] += 0.03 * bias_alpha
+        # Corners get micro bias (+0.01) for minimal structural support
+        exp_reshaped[:, corners, 1] += 0.01 * bias_alpha 
+        x_d_info["exp"] = exp_reshaped.reshape(1, -1)
 
         # [FIX] Expression Temporal Smoothing (EMA)
         # Prevents inhumanly fast mouth open/close cycles by adding inertia.

pyproject.toml

@@ -1,7 +1,7 @@
 [project]
 name = "aiia"
 description = "The Ultimate AI Audio/Video toolkit for ComfyUI. Features an enhanced Ditto (with optimizations that outperform official demos and other SOTA talking head models in lip-sync accuracy and natural motion), EchoMimic V3 & FLOAT, VibeVoice & CosyVoice 3.0 (Zero-Shot Voice Cloning), Multi-Role Podcast Generation, and a powerful Media Browser."
-version = "1.9.116"
+version = "1.9.117"
 license = {file = "LICENSE"}
 readme = "README.md"
 authors = [