v1.9.189: Predictive Pitch Bias & Volumetric Jaw Excitation. Implemented constant chin-tuck pressure and boosted lower lip gain to 1.2 to fix head-up drift and lazy mouth opening.

libs/Ditto/core/atomic_components/audio2motion.py

@@ -119,8 +119,8 @@ class Audio2Motion:
         self.brownian_momentum = np.zeros_like(self.kp_cond) # [v1.9.139] Postural inertia
         self.look_up_timer = 0 # [v1.9.141] Timer for anti-stall recovery
         self.is_recovering = False # [v1.9.155] Hysteresis state flag
-        self.target_bias_deg = 0.0 # [v1.9.162/163] For scope visibility
-        self.current_push = 0.0 # [v1.9.164] SAFETY: Initialize missing variable to prevent thread crash
+        self.target_bias_deg = -3.5 # [v1.9.189] Aesthetic Sweet Spot (Target Delta)
+        self.current_push = 0.0012 # [v1.9.189] Constant Chin-Tuck Pressure
         
         # [v1.9.170] Pure Photo Anchor (Reverted Neutralizer)
         self.photo_base_neutralizer = np.zeros_like(self.s_kp_cond)
@@ -248,15 +248,13 @@ class Audio2Motion:
             # [v1.9.164] Predictive Physics Push
             new_drift[0, 1:67] += self.current_push
             
-            # [v1.9.152] Absolute Postural Force (Boosted to 0.0030)
+            # [v1.9.189] Decoupled Axis Force (Chin-Tuck Mode)
             if self.is_recovering:
-                # [v1.9.158] Decoupled Axis Force
-                # High tension for Pitch (1:67), Soft tension for Yaw/Roll/T (67:202)
-                new_drift[0, 1:67] -= 0.0030 
-                new_drift[0, 67:202] -= 0.0010 # Softened Yaw/Roll pulse
+                # Add drift (push down) to recover from looking up too much
+                new_drift[0, 1:67] += 0.0035 
+                new_drift[0, 67:202] -= 0.0010 
                 if self.look_up_timer > 100:
-                    new_drift[0, 1:67] -= 0.0030 # Double impulse for Pitch
-            
+                    new_drift[0, 1:67] += 0.0035 
             self.brownian_momentum = self.brownian_momentum * 0.92 + new_drift
             self.brownian_pos += self.brownian_momentum
             
@@ -388,7 +386,7 @@ class Audio2Motion:
             
             if self.clip_idx % 20 == 0:
                  mode_s = "SPEECH" if is_talking else "IDLE"
-                 print(f"[v1.9.170 {mode_s}] Pressure: {pressure*100:.0f}% (Safety Net Active)")
+                 print(f"[v1.9.189 {mode_s}] Pressure: {pressure*100:.0f}% (Delta={delta_p:+.2f} Target={self.target_bias_deg:+.1f})")
         
         # [v1.9.156] Virtual Last Frame for Startup Stabilization
         # If this is the VERY first chunk, we treat the source photo as the "prev frame"

libs/Ditto/core/atomic_components/motion_stitch.py

@@ -533,8 +533,8 @@ class MotionStitch:
         self.fix_exp_a2 = (1 - _a1) + _a1 * _a2
         self.fix_exp_a3 = _a2
         
-        # [Debug v1.9.188] Verify Code Sync
-        print(f"[AIIA Debug] MotionStitch Setup: v1.9.188. LATEST VERSION LOADED.")
+        # [Debug v1.9.189] Verify Code Sync
+        print(f"[AIIA Debug] MotionStitch Setup: v1.9.189. LATEST VERSION LOADED.")
 
 
         if self.drive_eye and self.delta_eye_arr is not None:
@@ -701,9 +701,8 @@ class MotionStitch:
         # Only boost if the AI is actually trying to open the mouth (y > 0)
         mask = y_intent > 0.001
         if np.any(mask):
-            # Gain(y) = 1.0 + 0.8 * exp(-y / 0.04)
-            # High gain for small y, moves to 1.0 as y increases.
-            gain = 1.0 + 0.8 * np.exp(-y_intent[mask] / 0.04)
+            # [v1.9.189] Volumetric Jaw Excitation (Boosted to 1.2)
+            gain = 1.0 + 1.2 * np.exp(-y_intent[mask] / 0.04)
             exp_reshaped[:, lower_lip, 1][mask] *= gain
 
         # 2. Volumetric Mouth Micro-Motion (Breathing + Corners 7,8)

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 = "1.9.188"
+version = "1.9.189"
 license = {file = "LICENSE"}
 readme = "README.md"
 authors = [