diff --git a/src/libImaging/Filter.c b/src/libImaging/Filter.c
index 7360f9681..5fb9b060c 100644
--- a/src/libImaging/Filter.c
+++ b/src/libImaging/Filter.c
@@ -480,10 +480,60 @@ ImagingFilter5x5(Imaging imOut, Imaging im, const float* kernel,
             UINT8* in1 = (UINT8*) im->image[y+1];
             UINT8* in2 = (UINT8*) im->image[y+2];
             UINT8* out = (UINT8*) imOut->image[y];
+            __m128 kernelx0 = _mm_set_ps(kernel[15], kernel[10], kernel[5], kernel[0]);
+            __m128 kernel00 = _mm_loadu_ps(&kernel[0+1]);
+            __m128 kernel10 = _mm_loadu_ps(&kernel[5+1]);
+            __m128 kernel20 = _mm_loadu_ps(&kernel[10+1]);
+            __m128 kernel30 = _mm_loadu_ps(&kernel[15+1]);
+            __m128 kernel40 = _mm_loadu_ps(&kernel[20+1]);
+            __m128 kernelx1 = _mm_set_ps(kernel[19], kernel[14], kernel[9], kernel[4]);
+            __m128 kernel01 = _mm_loadu_ps(&kernel[0+0]);
+            __m128 kernel11 = _mm_loadu_ps(&kernel[5+0]);
+            __m128 kernel21 = _mm_loadu_ps(&kernel[10+0]);
+            __m128 kernel31 = _mm_loadu_ps(&kernel[15+0]);
+            __m128 kernel41 = _mm_loadu_ps(&kernel[20+0]);
 
             out[0] = in0[0];
             out[1] = in0[1];
-            for (x = 2; x < im->xsize-2; x++) {
+            x = 2;
+            for (; x < im->xsize-2-1; x += 2) {
+                __m128 ss0, ss1;
+                __m128i ssi0;
+                __m128 pix00, pix10, pix20, pix30, pix40;
+
+                MM_KERNEL1x5_LOAD(0, x-1);
+
+                ss0 = _mm_setzero_ps();
+                ss0 = _mm_add_ps(ss0, _mm_mul_ps(pix00, kernel00));
+                ss0 = _mm_add_ps(ss0, _mm_mul_ps(pix10, kernel10));
+                ss0 = _mm_add_ps(ss0, _mm_mul_ps(pix20, kernel20));
+                ss0 = _mm_add_ps(ss0, _mm_mul_ps(pix30, kernel30));
+                ss0 = _mm_add_ps(ss0, _mm_mul_ps(pix40, kernel40));
+                ss0 = _mm_add_ps(ss0, _mm_mul_ps(
+                    _mm_set_ps(in_1[x-2], in0[x-2], in1[x-2], in2[x-2]),
+                    kernelx0));
+
+                ss1 = _mm_setzero_ps();
+                ss1 = _mm_add_ps(ss1, _mm_mul_ps(pix00, kernel01));
+                ss1 = _mm_add_ps(ss1, _mm_mul_ps(pix10, kernel11));
+                ss1 = _mm_add_ps(ss1, _mm_mul_ps(pix20, kernel21));
+                ss1 = _mm_add_ps(ss1, _mm_mul_ps(pix30, kernel31));
+                ss1 = _mm_add_ps(ss1, _mm_mul_ps(pix40, kernel41));
+                ss1 = _mm_add_ps(ss1, _mm_mul_ps(
+                    _mm_set_ps(in_1[x+3], in0[x+3], in1[x+3], in2[x+3]),
+                    kernelx1));
+
+                ss0 = _mm_hadd_ps(ss0, ss1);
+                ss0 = _mm_add_ps(ss0, _mm_set_ps(
+                    offset, kernel[24] * in_2[x+3],
+                    offset, kernel[20] * in_2[x-2]));
+                ss0 = _mm_hadd_ps(ss0, ss0);
+                ssi0 = _mm_cvtps_epi32(ss0);
+                ssi0 = _mm_packs_epi32(ssi0, ssi0);
+                ssi0 = _mm_packus_epi16(ssi0, ssi0);
+                *((UINT32*) &out[x]) = _mm_cvtsi128_si32(ssi0);
+            }
+            for (; x < im->xsize-2; x++) {
                 float ss = offset;
                 ss += KERNEL1x5(in2, x, &kernel[0], 1);
                 ss += KERNEL1x5(in1, x, &kernel[5], 1);