From b0ce033002d3f9dd2dea41cf06b334e4b51b91e6 Mon Sep 17 00:00:00 2001
From: Hawk Lee <l.y@live.cn>
Date: Sat, 24 Jan 2026 20:07:04 +0800
Subject: [PATCH] v1.9.161: Absolute Neutral Anchor - Cancel photo tilt from
 Frame 0 to fix initial 8s baring

---
 .../Ditto/core/atomic_components/audio2motion.py | 16 ++++++++++++----
 pyproject.toml                                   |  2 +-
 2 files changed, 13 insertions(+), 5 deletions(-)

diff --git a/libs/Ditto/core/atomic_components/audio2motion.py b/libs/Ditto/core/atomic_components/audio2motion.py
index f93cfcf..5a2af0c 100644
--- a/libs/Ditto/core/atomic_components/audio2motion.py
+++ b/libs/Ditto/core/atomic_components/audio2motion.py
@@ -125,6 +125,13 @@ class Audio2Motion:
         p_s = e_s / e_s.sum()
         self.s_pitch_deg = np.sum(p_s * np.arange(66)) * 3 - 97.5
         print(f"[Postural Setup] Source Pitch Base: {self.s_pitch_deg:.2f}°")
+        
+        # [v1.9.161] Absolute Neutral Baseline (Zero-Drift Goal)
+        # We store an offset that cancelling out the source photo's postural bias.
+        # This allows us to anchor to a TRUE "Flat View" (0,0,0) from frame 0.
+        self.posture_neutralizer = np.zeros_like(self.s_kp_cond)
+        # Index 1:202 contains Scale(1), Pitch(66), Yaw(66), Roll(66), T(3)
+        self.posture_neutralizer[0, 1:202] = -self.s_kp_cond[0, 1:202]
 
     def _fuse(self, res_kp_seq, pred_kp_seq, override_alpha=None, step_len=None):
         ## ========================
@@ -333,14 +340,15 @@ class Audio2Motion:
         
         # [v1.9.160] STATIC POSTURE MODE
         # Based on validation, we now enforce high Pitch pressure ALWAYS to prevent drift.
-        # This eliminates the dependency on 'Hysteresis Recovery' and ensures the jaw stays stable.
+        # [v1.9.161] Neutral Baseline: Anchor is shifted to TRUE 0 (Neutral Center).
         if True:
-            # [v1.9.160] Constant 95% tension for Pitch (Vertical)
+            # Constant 95% tension for Pitch (Vertical)
             pressure = 0.95
             soft_p = 0.40 # Maintain soft-lock for Yaw/Roll to allow mouth projection
             
-            # Anchor is source + current Brownian offset
-            anchor_p = self.s_kp_cond[0, 1:202] + self.brownian_pos[0, 1:202]
+            # [v1.9.161] Anchor = Neutralized Source (0-Posture) + Brownian micro-drift
+            # This ensures we anchor to FLAT view, even if the photo is tilted.
+            anchor_p = (self.s_kp_cond + self.posture_neutralizer + self.brownian_pos)[0, 1:202]
             
             for f in range(pred_kp_seq.shape[1]):
                 # Absolute Pitch Locking (Halt baring distortion)
diff --git a/pyproject.toml b/pyproject.toml
index 29f4332..09406c3 100755
--- a/pyproject.toml
+++ b/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 = "1.9.160"
+version = "1.9.161"
 license = {file = "LICENSE"}
 readme = "README.md"
 authors = [