diff --git a/src/pixie/masks.nim b/src/pixie/masks.nim
index 7644bba..32d21b1 100644
--- a/src/pixie/masks.nim
+++ b/src/pixie/masks.nim
@@ -213,17 +213,15 @@ proc getValueSmooth*(mask: Mask, x, y: float32): uint8 {.raises: [].} =
 
 proc invert*(mask: Mask) {.raises: [].} =
   ## Inverts all of the values - creates a negative of the mask.
-  var i: int
-  when defined(amd64) and allowSimd:
-    let vec255 = mm_set1_epi8(255)
-    for _ in 0 ..< mask.data.len div 16:
-      var values = mm_loadu_si128(mask.data[i].addr)
-      values = mm_sub_epi8(vec255, values)
-      mm_storeu_si128(mask.data[i].addr, values)
-      i += 16
+  when allowSimd and compiles(invertImageSimd):
+    invertMaskSimd(
+      cast[ptr UncheckedArray[uint8]](mask.data[0].addr),
+      mask.data.len
+    )
+    return
 
-  for j in i ..< mask.data.len:
-    mask.data[j] = 255 - mask.data[j]
+  for i in 0 ..< mask.data.len:
+    mask.data[i] = 255 - mask.data[i]
 
 proc spread*(mask: Mask, spread: float32) {.raises: [PixieError].} =
   ## Grows the mask by spread.
diff --git a/src/pixie/simd.nim b/src/pixie/simd.nim
index 13453d1..9c100ec 100644
--- a/src/pixie/simd.nim
+++ b/src/pixie/simd.nim
@@ -255,3 +255,15 @@ when defined(amd64):
       data[i] = rgbx
 
     toPremultipliedAlphaSimd(cast[ptr UncheckedArray[uint32]](data), len)
+
+  proc invertMaskSimd*(data: ptr UncheckedArray[uint8], len: int) =
+    var i: int
+    let vec255 = mm_set1_epi8(255)
+    for _ in 0 ..< len div 16:
+      var values = mm_loadu_si128(data[i].addr)
+      values = mm_sub_epi8(vec255, values)
+      mm_storeu_si128(data[i].addr, values)
+      i += 16
+
+    for j in i ..< len:
+      data[j] = 255 - data[j]