nodes/havvk-ComfyUI_AIIA
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

v1.9.165: Fluid Vitality - Synthetic Gaussian Anchoring with lowered pressure (75% Pitch / 30% Global) to restore expressiveness

Browse Source
This commit is contained in:
Hawk Lee
2026-01-24 20:56:29 +08:00
parent ba476a4c6c
commit 4d91db892f
2 changed files with 52 additions and 33 deletions
Download Patch File Download Diff File Expand all files Collapse all files
libs/Ditto/core/atomic_components/audio2motion.py
+51 -32
View File
@@ -166,6 +166,18 @@ class Audio2Motion:
return res_kp_seq
def _generate_gaussian_pitch(self, target_deg, sigma=1.5):
# Ditto Pitch: 66 bins, 3 deg each, center at 32.5 (0 deg)
# target_deg = (idx * 3) - 97.5 => idx = (target_deg + 97.5) / 3
mu = (target_deg + 97.5) / 3.0
x = np.arange(66)
# We generate raw coefficients that will simulate the target degree
gaussian = np.exp(-((x - mu)**2) / (2 * sigma**2))
# Add tiny epsilon to avoid log(0) issues in model
gaussian = np.clip(gaussian, 1e-6, 1.0)
# Log-space coefficients for the categorical model
return np.log(gaussian).reshape(1, 66)
def _update_kp_cond(self, res_kp_seq, idx, step_len=0):
# [v1.9.144] Unconditional State Tracking
# We process these outside the fix_kp_cond branches to ensure monitoring never stops.
@@ -205,39 +217,45 @@ class Audio2Motion:
if self.is_recovering:
self.look_up_timer += step_len
# [v1.9.162] PREDICTIVE POSTURE STATE MACHINE
# We scan the VAD timeline to decide the character's "Readiness"
# [v1.9.165] FLUID POSTURE STATE MACHINE
# Decisions are based on Silence Frames (Universal) + VAD Lookahead (Optional)
target_bias_deg = 0.0 # Standard flat
is_currently_talking = self.silence_frames < 25
has_upcoming_speech = False
if self.vad_timeline is not None:
# Lookahead window: 2 seconds (50 frames)
lookahead = self.vad_timeline[idx : idx + 50]
is_currently_talking = self.silence_frames < 25
has_upcoming_speech = len(lookahead) > 0 and np.max(lookahead) > 0.1
if is_currently_talking or has_upcoming_speech:
# COMBAT STANCE: Tucked chin (-5 deg downward = -0.0020 impulse push)
target_bias_deg = +5.0
self.target_push = -0.0025 # Persistent negative impulse to lower head
else:
# IDLE SWAY: Natural breathing motion
cycle = np.sin(self.global_time * 0.05)
target_bias_deg = -3.0 + cycle * 5.0
self.target_push = 0.0 # No forced push, just natural sway
if is_currently_talking or has_upcoming_speech:
# COMBAT STANCE: Tucked chin for headroom
target_bias_deg = -5.0
else:
self.target_push = 0.0
target_bias_deg = 0.0
# Smooth Push Transition
self.current_push = self.current_push * 0.90 + self.target_push * 0.10
# IDLE SWAY: Natural breathing
cycle = np.sin(self.global_time * 0.05)
target_bias_deg = +2.0 + cycle * 5.0 # Slightly down/neutral sway
# [v1.9.165] Synthetic Anchor Generation
# Instead of photo bias, we anchor to a mathematical IDEAL distribution.
self.target_bias_deg = target_bias_deg
synthetic_pitch = self._generate_gaussian_pitch(target_bias_deg)
# Apply to anchor target
self.target_neutralizer = self.photo_base_neutralizer.copy()
# Offset 1:67 is Pitch. We overwrite with synthetic distribution.
# Since photo_base_neutralizer[1:67] is -s_kp[1:67],
# photo_base + s_kp + synthetic = synthetic.
self.target_neutralizer[0, 1:67] = -self.s_kp_cond[0, 1:67] + synthetic_pitch[0]
# Smooth Transition (0.05 EMA = ~1s glide)
self.current_neutralizer = self.current_neutralizer * 0.95 + self.target_neutralizer * 0.05
tag = "[HEARTBEAT]" if not self.is_recovering else "[RECOVERY]"
if self.vad_timeline is not None and has_upcoming_speech and not is_currently_talking:
if has_upcoming_speech and not is_currently_talking:
tag = "[ANTICIPATION]"
print(f"{tag} Frame {idx:04d} | Delta={delta_p:+.2f}° | Bias={self.target_bias_deg:+.1f}° | Silence={self.silence_frames:03d}")
print(f"{tag} Frame {idx:04d} | Delta={delta_p:+.2f}° | Goal={self.target_bias_deg:+.1f}° | Silence={self.silence_frames:03d}")
# 5. Postural Auto-Correction Logic [REPLACED by v1.9.160 STATIC + v1.9.162 PREDICTIVE]
@@ -381,25 +399,26 @@ class Audio2Motion:
pred_kp_seq = self.lmdm(self.kp_cond, aud_cond, self.sampling_timesteps)
# [v1.9.160/162/164] STATIC/PREDICTIVE POSTURE MODE
# We now enforce high Pitch pressure ALWAYS to anchor to our target baseline.
# [v1.9.165] FLUID POSTURE MODE
# We now use "Mathematical Gravity" to pull the head into the safe zone.
if True:
# Constant 95% tension for Pitch (Vertical)
pressure = 0.95
soft_p = 0.40 # Maintain soft-lock for Yaw/Roll to allow mouth projection
# [v1.9.165] Fluid Pressure: 75% Pitch / 30% Yaw-Roll
# This restores expressiveness while maintaining vertical stability.
pressure = 0.75
soft_p = 0.30
# [v1.9.164] Stable Anchor = Photo + Drift-Adjusted Offset
anchor_p = (self.s_kp_cond + self.brownian_pos)[0, 1:202]
# Stable Anchor = Photo + Synthetic Offset + Brownian Drift
anchor_p = (self.s_kp_cond + self.current_neutralizer + self.brownian_pos)[0, 1:202]
for f in range(pred_kp_seq.shape[1]):
# Absolute Pitch Locking (Halt baring distortion)
pred_kp_seq[0, f, 1:67] = pred_kp_seq[0, f, 1:67] * (1.0 - pressure) + anchor_p[0:66] * pressure
# Soft Rotational Locking (Organic freedom)
# Soft Rotational Locking (Organic movement)
pred_kp_seq[0, f, 67:202] = pred_kp_seq[0, f, 67:202] * (1.0 - soft_p) + anchor_p[66:201] * soft_p
if self.clip_idx % 20 == 0:
mode_s = "SPEECH" if self.target_bias_deg > 2.0 else "IDLE"
print(f"[v1.9.162 {mode_s}] Anchor: {self.target_bias_deg:+.1f}° goal.")
mode_s = "SPEECH" if self.target_bias_deg < -2.0 else "IDLE"
print(f"[v1.9.165 {mode_s}] Fluid Gravity: {self.target_bias_deg:+.1f}° goal.")
# [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"
pyproject.toml
+1 -1
View File
@@ -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 = "1.9.164"
version = "1.9.165"
license = {file = "LICENSE"}
readme = "README.md"
authors = [