Update genmake_voxel_engine.py

genmake_voxel_engine.py

@@ -1,9 +1,6 @@
 import numpy as np
 import cv2
 
-# ----------------------------
-# GenMake Core Tunables
-# ----------------------------
 DEFAULT_VOXEL_MM = 0.024
 DEFAULT_DETAIL_BOOST = 0.5  
 BAYER_STRENGTH = 0.75
@@ -44,6 +41,15 @@ def build_importance_map(rgb_bgr: np.ndarray, depth01: np.ndarray, fg_mask01: np
     imp = (sharp * 0.70) + (edges * 0.20) + (depth_edges * 0.10)
     imp = _norm01(imp)
 
+    # --- UPGRADE 2: Face Priority Boost ---
+    h, w = rgb_bgr.shape[:2]
+    cx, cy = w // 2, h // 2
+    Y, X = np.ogrid[:h, :w]
+    dist_sq = (X - cx)**2 + (Y - cy)**2
+    max_dist_sq = cx**2 + cy**2 + 1e-5
+    face_priority = 1.0 - (dist_sq / max_dist_sq)
+    imp = imp * (0.7 + 0.3 * face_priority).astype(np.float32)
+
     if fg_mask01 is not None:
         m = _to_mask01(fg_mask01)
         m = cv2.GaussianBlur(m, (0, 0), 1.2)
@@ -98,22 +104,23 @@ def sample_points_from_volume(V: tuple, total_points: int, voxel_mm: float, sx_m
     prob = np.clip(imp_grid * scale, 0.0, 1.0).astype(np.float32)
 
     score = prob - (bayer * BAYER_STRENGTH)
-    flat = score.reshape(-1)
+    flat = score.flatten()
 
     if not diorama_mode:
-        flat_mask = mask_grid.reshape(-1)
+        flat_mask = mask_grid.flatten()
         flat[flat_mask < 0.5] = -9999.0
 
     valid_point_count = np.sum(flat > -9000.0)
     actual_target = min(int(total_points), int(valid_point_count))
     
-    k = int(max(1, min(flat.size, actual_target)))
-    thr = np.partition(flat, -k)[-k]
-    
-    if not diorama_mode:
-        final_mask = (score >= thr) & (score > -9000.0) & (mask_grid >= 0.5)
+    # --- UPGRADE 3: Deterministic Point Targeting ---
+    if actual_target > 0:
+        top_idx = np.argpartition(flat, -actual_target)[-actual_target:]
+        final_mask_flat = np.zeros_like(flat, dtype=bool)
+        final_mask_flat[top_idx] = True
+        final_mask = final_mask_flat.reshape(grid_h, grid_w)
     else:
-        final_mask = (score >= thr)
+        final_mask = np.zeros((grid_h, grid_w), dtype=bool)
 
     ys, xs = np.nonzero(final_mask)
     if xs.size == 0:
@@ -122,7 +129,8 @@ def sample_points_from_volume(V: tuple, total_points: int, voxel_mm: float, sx_m
     xx = (xs.astype(np.float32) / (grid_w - 1) - 0.5) * float(sx_mm)
     yy = (0.5 - ys.astype(np.float32) / (grid_h - 1)) * float(sy_mm)
 
-    depth_span_mm = float(np.clip(float(sx_mm) * depth_vol, 8.0, float(sz_mm) * 0.90))
+    # --- UPGRADE 4: Z-Axis Plaque Fix ---
+    depth_span_mm = min(float(sz_mm) * depth_vol, float(sz_mm) * 0.90)
 
     zbase = dep_grid[ys, xs].astype(np.float32)
     p5 = float(np.percentile(zbase, 5))
@@ -141,8 +149,9 @@ def sample_points_from_volume(V: tuple, total_points: int, voxel_mm: float, sx_m
 
     zz -= float(zz.min())
 
+    # --- UPGRADE 5: UV-Optimized Micro Jitter ---
     rng = np.random.default_rng(int(seed))
-    xy_j = float(voxel_mm) * 0.15
+    xy_j = float(voxel_mm) * 0.08 
     z_j = 0.05 
     
     xx = (xx + (rng.random(xx.shape, dtype=np.float32) - 0.5) * xy_j).astype(np.float32)