diff --git a/cuburn/code/filtering.py b/cuburn/code/filtering.py
index 10515bc..702c15a 100644
--- a/cuburn/code/filtering.py
+++ b/cuburn/code/filtering.py
@@ -8,10 +8,10 @@ class ColorClip(HunkOCode):
     defs_tmpl = Template('''
 __global__
 void colorclip(float4 *pixbuf, float gamma, float vibrancy, float highpow,
-               float linrange, float lingam, float3 bkgd) {
-    // TODO: test if over an edge of the framebuffer - currently gutters are
-    // used and up to 256 pixels are ignored, which breaks when width<256
-    int i = (gridDim.x * blockIdx.y + blockIdx.x) * blockDim.x + threadIdx.x;
+               float linrange, float lingam, float3 bkgd, int fbsize) {
+    int i = gtid();
+    if (i >= fbsize) return;
+
     float4 pix = pixbuf[i];
 
     if (pix.w <= 0) {
diff --git a/cuburn/render.py b/cuburn/render.py
index d79659d..28dd6bd 100644
--- a/cuburn/render.py
+++ b/cuburn/render.py
@@ -338,7 +338,7 @@ class Animation(object):
 
             color_fun = self.mod.get_function("colorclip")
             blocks = int(np.ceil(np.sqrt(nbins / 256)))
-            color_fun(d_out, gam, vib, hipow, lin, lingam, bkgd,
+            color_fun(d_out, gam, vib, hipow, lin, lingam, bkgd, np.int32(nbins),
                       block=(256, 1, 1), grid=(blocks, blocks),
                       stream=filt_stream)
             cuda.memcpy_dtoh_async(h_out, d_out, filt_stream)