diff --git a/src/pixie/images.nim b/src/pixie/images.nim
index a5c2e6f..c2fc65d 100644
--- a/src/pixie/images.nim
+++ b/src/pixie/images.nim
@@ -329,6 +329,46 @@ proc applyOpacity*(target: Image | Mask, opacity: float32) =
     for j in i ..< target.data.len:
       target.data[j] = ((target.data[j] * opacity) div 255).uint8
 
+proc invert*(target: Image | Mask) =
+  ## Inverts all of the colors and alpha.
+  var i: int
+  when defined(amd64) and not defined(pixieNoSimd):
+    let v255 = mm_set1_epi8(cast[int8](255))
+
+    when type(target) is Image:
+      let byteLen = target.data.len * 4
+    else:
+      let byteLen = target.data.len
+
+    for _ in countup(0, byteLen - 16, 16):
+      when type(target) is Image:
+        let index = i div 4
+      else:
+        let index = i
+
+      var values = mm_loadu_si128(target.data[index].addr)
+      values = mm_sub_epi8(v255, values)
+      mm_storeu_si128(target.data[index].addr, values)
+
+      i += 16
+
+  when type(target) is Image:
+    for j in i div 4 ..< target.data.len:
+      var rgba = target.data[j]
+      rgba.r = 255 - rgba.r
+      rgba.g = 255 - rgba.g
+      rgba.b = 255 - rgba.b
+      rgba.a = 255 - rgba.a
+      target.data[j] = rgba
+
+    # Inverting rgba(50, 100, 150, 200) becomes rgba(205, 155, 105, 55). This
+    # is not a valid premultiplied alpha color.
+    # We need to convert back to premultiplied alpha after inverting.
+    target.toPremultipliedAlpha()
+  else:
+    for j in i ..< target.data.len:
+      target.data[j] = (255 - target.data[j]).uint8
+
 proc getRgbaSmooth*(image: Image, x, y: float32): ColorRGBA =
   let
     minX = floor(x)
@@ -439,24 +479,6 @@ proc draw*(
 when defined(release):
   {.pop.}
 
-proc invert*(image: Image) =
-  ## Inverts all of the colors and alpha.
-  var i: int
-  when defined(amd64) and not defined(pixieNoSimd):
-    let vec255 = mm_set1_epi8(cast[int8](255))
-    while i < image.data.len - 4:
-      var m = mm_loadu_si128(image.data[i].addr)
-      m = mm_sub_epi8(vec255, m)
-      mm_storeu_si128(image.data[i].addr, m)
-      i += 4
-  for j in i ..< image.data.len:
-    var rgba = image.data[j]
-    rgba.r = 255 - rgba.r
-    rgba.g = 255 - rgba.g
-    rgba.b = 255 - rgba.b
-    rgba.a = 255 - rgba.a
-    image.data[j] = rgba
-
 proc gaussianLookup(radius: int): seq[float32] =
   ## Compute lookup table for 1d Gaussian kernel.
   result.setLen(radius * 2 + 1)
diff --git a/tests/benchmark_masks.nim b/tests/benchmark_masks.nim
index 5289fdf..3693d7c 100644
--- a/tests/benchmark_masks.nim
+++ b/tests/benchmark_masks.nim
@@ -13,5 +13,10 @@ timeIt "minifyBy2":
 
 reset()
 
+timeIt "invert":
+  mask.invert()
+
+reset()
+
 timeIt "applyOpacity":
   mask.applyOpacity(0.5)
diff --git a/tests/test_images.nim b/tests/test_images.nim
index cccff3b..0a56c7b 100644
--- a/tests/test_images.nim
+++ b/tests/test_images.nim
@@ -98,3 +98,9 @@ block:
     a = readImage("tests/images/flipped1.png")
     b = a.minifyBy2(2)
   b.writeFile("tests/images/minifiedBy4.png")
+
+block:
+  let a = newImage(100, 100)
+  a.fill(rgba(50, 100, 150, 200))
+  a.invert()
+  doAssert a[0, 0] == rgba(44, 33, 22, 55)
diff --git a/tests/test_masks.nim b/tests/test_masks.nim
index 02b5aaa..43ffd2e 100644
--- a/tests/test_masks.nim
+++ b/tests/test_masks.nim
@@ -7,6 +7,12 @@ block:
   doAssert mask[0, 0] == 100
   doAssert mask[88, 88] == 100
 
+block:
+  let mask = newMask(100, 100)
+  mask.fill(200)
+  mask.invert()
+  doAssert mask[0, 0] == 55
+
 block:
   let
     mask = newMask(100, 100)