Add heightmaps to burley tiling

Also add --homo flag to gaussianize. This "homogenizes" the input by
removing low frequency energy from the DFT of the image luma.

1 files changed, 61 insertions, 0 deletions

diff --git a/Scripts/gaussianize b/Scripts/gaussianize
index 2ca915d..620e05d 100755
--- a/Scripts/gaussianize
+++ b/Scripts/gaussianize
@@ -72,6 +72,58 @@ def save_image(image: np.ndarray, output_path: Path):
         raise ValueError(f"Unsupported output format '{output_path.suffix}'. Use .exr or .png.")
 
 
+def rgb_to_ycrcb(image: np.ndarray) -> np.ndarray:
+    """Convert RGB in [0, 1] to full-range YCrCb in [0, 1]."""
+    rgb = np.clip(image, 0.0, 1.0)
+    r = rgb[:, :, 0]
+    g = rgb[:, :, 1]
+    b = rgb[:, :, 2]
+
+    y = 0.299 * r + 0.587 * g + 0.114 * b
+    cb = 0.5 + (b - y) * 0.564
+    cr = 0.5 + (r - y) * 0.713
+    return np.stack((y, cr, cb), axis=-1)
+
+
+def ycrcb_to_rgb(image: np.ndarray) -> np.ndarray:
+    """Convert full-range YCrCb in [0, 1] back to RGB in [0, 1]."""
+    ycrcb = np.asarray(image, dtype=np.float64)
+    y = ycrcb[:, :, 0]
+    cr = ycrcb[:, :, 1] - 0.5
+    cb = ycrcb[:, :, 2] - 0.5
+
+    r = y + 1.402 * cr
+    g = y - 0.714136 * cr - 0.344136 * cb
+    b = y + 1.772 * cb
+    return np.clip(np.stack((r, g, b), axis=-1), 0.0, 1.0)
+
+
+def subtract_low_frequencies(channel: np.ndarray, cutoff_cycles: float = 8.0) -> np.ndarray:
+    """Subtract a smooth periodic low-frequency component from a scalar image."""
+    if cutoff_cycles <= 0.0:
+        raise ValueError(f"cutoff_cycles must be > 0, got {cutoff_cycles}")
+
+    channel = np.asarray(channel, dtype=np.float64)
+    height, width = channel.shape
+
+    fy = np.fft.fftfreq(height)
+    fx = np.fft.fftfreq(width)
+    radius = np.sqrt((fy[:, np.newaxis] * height) ** 2 + (fx[np.newaxis, :] * width) ** 2)
+
+    low_mask = np.exp(-((radius / cutoff_cycles) ** 2))
+    low_pass = np.fft.ifft2(np.fft.fft2(channel) * low_mask).real
+
+    # Remove only the spatial trend and keep the original average luminance.
+    return channel - low_pass + float(np.mean(low_pass))
+
+
+def homogenize_luminance(image: np.ndarray, cutoff_cycles: float = 8.0) -> np.ndarray:
+    """Reduce very-low-frequency luminance variation while preserving chroma."""
+    ycrcb = rgb_to_ycrcb(image)
+    ycrcb[:, :, 0] = subtract_low_frequencies(ycrcb[:, :, 0], cutoff_cycles=cutoff_cycles)
+    return ycrcb_to_rgb(ycrcb)
+
+
 class TruncatedGaussian:
     """Truncated Gaussian distribution for histogram transformation."""
 
@@ -450,6 +502,11 @@ def main():
         action="store_true",
         help="Print detailed progress information"
     )
+    parser.add_argument(
+        "--homo",
+        action="store_true",
+        help="Homogenize luminance before gaussianizing by removing very-low-frequency Y in YCrCb space"
+    )
 
     args = parser.parse_args()
 
@@ -467,6 +524,10 @@ def main():
         image = load_image(args.input)
         quantization_step = 0.0 if args.input.suffix.lower() == ".exr" else (1.0 / 255.0)
 
+        if args.homo:
+            print("Homogenizing luminance in YCrCb space...")
+            image = homogenize_luminance(image)
+
         # Gaussianize the texture
         print("Applying per-channel Gaussianization...")
         gaussianized, inverse_luts = gaussianize_texture(