diff --git a/src/libImaging/Filter.c b/src/libImaging/Filter.c
index efec61657..1bc882e95 100644
--- a/src/libImaging/Filter.c
+++ b/src/libImaging/Filter.c
@@ -244,6 +244,57 @@ ImagingFilter3x3(Imaging imOut, Imaging im, const float* kernel,
             UINT32* in0 = (UINT32*) im->image[y];
             UINT32* in1 = (UINT32*) im->image[y+1];
             UINT32* out = (UINT32*) imOut->image[y];
+#if defined(__AVX2__)
+            __m256 kernel00 = _mm256_insertf128_ps(
+                _mm256_set1_ps(kernel[0+0]),
+                _mm_set1_ps(kernel[0+1]), 1);
+            __m256 kernel01 = _mm256_castps128_ps256(_mm_set1_ps(kernel[0+2]));
+            __m256 kernel10 = _mm256_insertf128_ps(
+                _mm256_set1_ps(kernel[3+0]),
+                _mm_set1_ps(kernel[3+1]), 1);
+            __m256 kernel11 = _mm256_castps128_ps256(_mm_set1_ps(kernel[3+2]));
+            __m256 kernel20 = _mm256_insertf128_ps(
+                _mm256_set1_ps(kernel[6+0]),
+                _mm_set1_ps(kernel[6+1]), 1);
+            __m256 kernel21 = _mm256_castps128_ps256(_mm_set1_ps(kernel[6+2]));
+            __m256 pix00, pix10, pix20;
+            __m256 pix01, pix11, pix21;
+
+            out[0] = in0[0];
+            x = 1;
+            pix00 = _mm256_cvtepi32_ps(_mm256_cvtepu8_epi32(*(__m128i*) &in1[0]));
+            pix10 = _mm256_cvtepi32_ps(_mm256_cvtepu8_epi32(*(__m128i*) &in0[0]));
+            pix20 = _mm256_cvtepi32_ps(_mm256_cvtepu8_epi32(*(__m128i*) &in_1[0]));
+            for (; x < im->xsize-1; x += 1) {
+                __m256 ss;
+                __m128i ssi;
+
+                pix01 = _mm256_cvtepi32_ps(_mm256_cvtepu8_epi32(*(__m128i*) &in1[x+1]));
+                pix11 = _mm256_cvtepi32_ps(_mm256_cvtepu8_epi32(*(__m128i*) &in0[x+1]));
+                pix21 = _mm256_cvtepi32_ps(_mm256_cvtepu8_epi32(*(__m128i*) &in_1[x+1]));
+
+                ss = _mm256_set1_ps(offset);
+                ss = _mm256_add_ps(ss, _mm256_mul_ps(pix00, kernel00));
+                ss = _mm256_add_ps(ss, _mm256_mul_ps(pix01, kernel01));
+                ss = _mm256_add_ps(ss, _mm256_mul_ps(pix10, kernel10));
+                ss = _mm256_add_ps(ss, _mm256_mul_ps(pix11, kernel11));
+                ss = _mm256_add_ps(ss, _mm256_mul_ps(pix20, kernel20));
+                ss = _mm256_add_ps(ss, _mm256_mul_ps(pix21, kernel21));
+
+                ssi = _mm_cvtps_epi32(_mm_add_ps(
+                    _mm256_extractf128_ps(ss, 0),
+                    _mm256_extractf128_ps(ss, 1)
+                ));
+                ssi = _mm_packs_epi32(ssi, ssi);
+                ssi = _mm_packus_epi16(ssi, ssi);
+                out[x] = _mm_cvtsi128_si32(ssi);
+
+                pix00 = _mm256_permute2f128_ps(pix00, pix01, 0x21);
+                pix10 = _mm256_permute2f128_ps(pix10, pix11, 0x21);
+                pix20 = _mm256_permute2f128_ps(pix20, pix21, 0x21);
+            }
+            out[x] = in0[x];
+#else
             __m128 pix00, pix10, pix20;
             __m128 pix01, pix11, pix21;
             __m128 pix02, pix12, pix22;
@@ -297,6 +348,7 @@ ImagingFilter3x3(Imaging imOut, Imaging im, const float* kernel,
                 out[x] = _mm_cvtsi128_si32(ssi0);
             }
             out[x] = in0[x];
+#endif
         }
     }
     memcpy(imOut->image[y], im->image[y], im->linesize);