Upload ldq_perceptual_layer.py with huggingface_hub

ldq_perceptual_layer.py

@@ -0,0 +1,63 @@
+#!/usr/bin/env python3
+"""
+LDQ Perceptual Feature Translation Layer
+Professional mixing effects driven by image statistics
+"""
+
+import numpy as np
+import math
+from scipy import signal
+from typing import Dict, Any
+
+def apply_perceptual_mixing(audio: np.ndarray, features: Dict[str, Any], sr: int) -> np.ndarray:
+    """
+    Apply multiband saturation, sidechain compression, stereo width, and reverb.
+    Returns polished audio.
+    """
+    if audio.ndim == 1:
+        audio = audio[:, np.newaxis]
+    
+    # 1. Multiband saturation based on contrast (reduced for less buzz)
+    contrast = features.get("contrast", 0.15)
+    sat_amount = contrast * 1.5  # Reduced from 3.0
+    
+    # Soft saturation
+    if sat_amount > 0:
+        audio = np.tanh(audio * (1 + sat_amount * 0.5)) / np.tanh(1 + sat_amount * 0.5)
+    
+    # 2. Sidechain compression based on hue
+    hue = features.get("average_hue", 180)
+    # Redder hue = faster release
+    release_time = 0.1 + (hue / 360) * 0.3  # 0.1-0.4 seconds
+    
+    # Gentle gain reduction based on edge density
+    edge_density = features.get("edge_density", 0.02)
+    reduction = 1.0 - (edge_density * 1.0)  # Less reduction
+    reduction = max(0.7, reduction)
+    audio *= reduction
+    
+    # 3. Stereo width based on keypoint distribution
+    chaos = features.get("chaos_keypoints", 0)
+    width = 1.0 + (chaos / 1000) * 0.3  # Less width
+    if audio.shape[1] == 2:
+        mid = (audio[:, 0] + audio[:, 1]) / 2
+        side = (audio[:, 0] - audio[:, 1]) / 2
+        audio[:, 0] = mid + side * width
+        audio[:, 1] = mid - side * width
+    
+    # 4. Reverb based on color variance (reduced)
+    colorfulness = features.get("colorfulness", 0.1)
+    reverb_decay = 0.3 + colorfulness * 3.0  # 0.3-3.3 seconds (shorter)
+    
+    if reverb_decay > 0.8:
+        delay_samples = int(sr * 0.03)  # 30ms (shorter)
+        if len(audio) > delay_samples:
+            reverb = np.roll(audio, delay_samples, axis=0)
+            reverb *= 0.2  # Less reverb
+            audio += reverb * min(0.6, reverb_decay / 4.0)
+    
+    # 5. Gentle EQ: cut above 5 kHz to reduce buzz
+    sos = signal.butter(2, 5000, btype='low', fs=sr, output='sos')
+    audio = signal.sosfilt(sos, audio, axis=0)
+    
+    return audio