diff --git a/pipi/filter/blur.c b/pipi/filter/blur.c
index 5270124..60247f0 100644
--- a/pipi/filter/blur.c
+++ b/pipi/filter/blur.c
@@ -27,6 +27,11 @@
 #include "pipi.h"
 #include "pipi_internals.h"
 
+#if !defined TEMPLATE_FILE /* This file uses the template system */
+#define TEMPLATE_FLAGS SET_FLAG_GRAY
+#define TEMPLATE_FILE "filter/blur.c"
+#include "pipi_template.h"
+
 /* Any standard deviation below this value will be rounded up, in order
  * to avoid ridiculously low values. exp(-1/(2*0.2*0.2)) is < 10^-5 so
  * there is little chance that any value below 0.2 will be useful. */
@@ -124,60 +129,64 @@ pipi_image_t *pipi_box_blur(pipi_image_t *src, int size)
     return pipi_box_blur_ext(src, size, size);
 }
 
-/* FIXME: split this into templates for wrap-around and proper gray support */
 pipi_image_t *pipi_box_blur_ext(pipi_image_t *src, int m, int n)
+{
+    if(src->last_modified == PIPI_PIXELS_Y_F)
+        return boxblur_gray(src, m, n);
+
+    return boxblur(src, m, n);
+}
+
+#else /* XXX: the following functions use the template system */
+
+/* FIXME: FLAG_WRAP */
+static pipi_image_t *SUFFIX(boxblur)(pipi_image_t *src, int m, int n)
 {
     pipi_image_t *dst;
     pipi_pixels_t *srcp, *dstp;
     float *srcdata, *dstdata;
     double *acc;
-    int x, y, w, h, i, j, size, gray;
+    int x, y, w, h, i, j, size;
 
     w = src->w;
     h = src->h;
     size = (2 * m + 1) * (2 * n + 1);
 
-    gray = (src->last_modified == PIPI_PIXELS_Y_F);
-
-    srcp = gray ? pipi_getpixels(src, PIPI_PIXELS_Y_F)
-                : pipi_getpixels(src, PIPI_PIXELS_RGBA_F);
+    srcp = FLAG_GRAY ? pipi_getpixels(src, PIPI_PIXELS_Y_F)
+                     : pipi_getpixels(src, PIPI_PIXELS_RGBA_F);
     srcdata = (float *)srcp->pixels;
 
     dst = pipi_new(w, h);
-    dstp = gray ? pipi_getpixels(dst, PIPI_PIXELS_Y_F)
-                : pipi_getpixels(dst, PIPI_PIXELS_RGBA_F);
+    dstp = FLAG_GRAY ? pipi_getpixels(dst, PIPI_PIXELS_Y_F)
+                     : pipi_getpixels(dst, PIPI_PIXELS_RGBA_F);
     dstdata = (float *)dstp->pixels;
 
-    acc = malloc(w * (gray ? 1 : 4) * sizeof(double));
+    acc = malloc(w * (FLAG_GRAY ? 1 : 4) * sizeof(double));
 
     /* Step 1: fill the accumulator */
     for(x = 0; x < w; x++)
     {
-        if(gray)
-        {
-            double t = 0.;
-
-            for(j = -n; j <= n; j++)
-            {
-                int j2 = (j < 0) ? h - 1 - ((-j - 1) % h) : j % h;
-                t += srcdata[j2 * w + x];
-            }
+        double r = 0., g = 0., b = 0., a = 0.;
+        double t = 0.;
 
-            acc[x] = t;
-        }
-        else
+        for(j = -n; j <= n; j++)
         {
-            double r = 0., g = 0., b = 0., a = 0.;
-
-            for(j = -n; j <= n; j++)
+            int j2 = (j < 0) ? h - 1 - ((-j - 1) % h) : j % h;
+            if(FLAG_GRAY)
+                t += srcdata[j2 * w + x];
+            else
             {
-                int j2 = (j < 0) ? h - 1 - ((-j - 1) % h) : j % h;
                 r += srcdata[4 * (j2 * w + x)];
                 g += srcdata[4 * (j2 * w + x) + 1];
                 b += srcdata[4 * (j2 * w + x) + 2];
                 a += srcdata[4 * (j2 * w + x) + 3];
             }
+        }
 
+        if(FLAG_GRAY)
+            acc[x] = t;
+        else
+        {
             acc[4 * x] = r;
             acc[4 * x + 1] = g;
             acc[4 * x + 2] = b;
@@ -192,19 +201,14 @@ pipi_image_t *pipi_box_blur_ext(pipi_image_t *src, int m, int n)
         double t = 0.;
 
         /* 2.1: compute the first pixel */
-        if(gray)
+        for(i = -m; i <= m; i++)
         {
-            for(i = -m; i <= m; i++)
-            {
-                int i2 = (i < 0) ? w - 1 - ((-i - 1) % w) : i % w;
+            int i2 = (i < 0) ? w - 1 - ((-i - 1) % w) : i % w;
+
+            if(FLAG_GRAY)
                 t += acc[i2];
-            }
-        }
-        else
-        {
-            for(i = -m; i <= m; i++)
+            else
             {
-                int i2 = (i < 0) ? w - 1 - ((-i - 1) % w) : i % w;
                 r += acc[4 * i2];
                 g += acc[4 * i2 + 1];
                 b += acc[4 * i2 + 2];
@@ -217,7 +221,7 @@ pipi_image_t *pipi_box_blur_ext(pipi_image_t *src, int m, int n)
         {
             int u, u2, v, v2;
 
-            if(gray)
+            if(FLAG_GRAY)
             {
                 dstdata[y * w + x] = t / size;
             }
@@ -233,7 +237,7 @@ pipi_image_t *pipi_box_blur_ext(pipi_image_t *src, int m, int n)
             u2 = (u < 0) ? w - 1 - ((-u - 1) % w) : u % w;
             v = x + m + 1;
             v2 = (v < 0) ? w - 1 - ((-v - 1) % w) : v % w;
-            if(gray)
+            if(FLAG_GRAY)
             {
                 t = t - acc[u2] + acc[v2];
             }
@@ -255,7 +259,7 @@ pipi_image_t *pipi_box_blur_ext(pipi_image_t *src, int m, int n)
             u2 = (u < 0) ? w - 1 - ((-u - 1) % w) : u % w;
             v = y + n + 1;
             v2 = (v < 0) ? w - 1 - ((-v - 1) % w) : v % w;
-            if(gray)
+            if(FLAG_GRAY)
             {
                 acc[x] += srcdata[v2 * w + x] - srcdata[u2 * w + x];
             }
@@ -277,3 +281,5 @@ pipi_image_t *pipi_box_blur_ext(pipi_image_t *src, int m, int n)
     return dst;
 }
 
+#endif
+