diff --git a/pipi/filter/blur.c b/pipi/filter/blur.c
index 3c5fc9f..5cc64a1 100644
--- a/pipi/filter/blur.c
+++ b/pipi/filter/blur.c
@@ -33,7 +33,7 @@ pipi_image_t *pipi_gaussian_blur(pipi_image_t *src, float radius)
     pipi_pixels_t *srcp, *dstp;
     float *srcdata, *dstdata;
     double *kernel;
-    double K, L, t = 0.;
+    double K, L;
     int x, y, i, j, w, h, kr, kw;
 
     w = src->w;
@@ -54,32 +54,39 @@ pipi_image_t *pipi_gaussian_blur(pipi_image_t *src, float radius)
     kernel = malloc(kw * kw * sizeof(double));
     for(j = -kr; j <= kr; j++)
         for(i = -kr; i <= kr; i++)
-            t += kernel[(j + kr) * kw + (i + kr)]
-                     = exp(L * (i * i + j * j)) * K;
-
-    for(j = -kr; j <= kr; j++)
-        for(i = -kr; i <= kr; i++)
-            kernel[(j + kr) * kw + (i + kr)] /= t;
+            kernel[(j + kr) * kw + (i + kr)] = exp(L * (i * i + j * j)) * K;
 
     for(y = 0; y < h; y++)
     {
         for(x = 0; x < w; x++)
         {
-            double R = 0., G = 0., B = 0.;
+            double R = 0., G = 0., B = 0., t = 0.;
+            int x2, y2;
 
             for(j = -kr; j <= kr; j++)
+            {
+                y2 = y + j;
+                if(y2 < 0) y2 = 0;
+                else if(y2 >= h) y2 = h - 1;
+
                 for(i = -kr; i <= kr; i++)
                 {
                     double f = kernel[(j + kr) * kw + (i + kr)];
 
-                    R += f * srcdata[((y + j) * w + x + i) * 4];
-                    G += f * srcdata[((y + j) * w + x + i) * 4 + 1];
-                    B += f * srcdata[((y + j) * w + x + i) * 4 + 2];
+                    x2 = x + i;
+                    if(x2 < 0) x2 = 0;
+                    else if(x2 >= w) x2 = w - 1;
+
+                    R += f * srcdata[(y2 * w + x2) * 4];
+                    G += f * srcdata[(y2 * w + x2) * 4 + 1];
+                    B += f * srcdata[(y2 * w + x2) * 4 + 2];
+                    t += f;
                 }
+            }
 
-            dstdata[(y * w + x) * 4] = R;
-            dstdata[(y * w + x) * 4 + 1] = G;
-            dstdata[(y * w + x) * 4 + 2] = B;
+            dstdata[(y * w + x) * 4] = R / t;
+            dstdata[(y * w + x) * 4 + 1] = G / t;
+            dstdata[(y * w + x) * 4 + 2] = B / t;
         }
     }