diff --git a/Tests/images/imagedraw_ellipse_L.png b/Tests/images/imagedraw_ellipse_L.png
index e47e6e441..d5959cc08 100644
Binary files a/Tests/images/imagedraw_ellipse_L.png and b/Tests/images/imagedraw_ellipse_L.png differ
diff --git a/Tests/images/imagedraw_ellipse_RGB.png b/Tests/images/imagedraw_ellipse_RGB.png
index b52b12802..1d310ce11 100644
Binary files a/Tests/images/imagedraw_ellipse_RGB.png and b/Tests/images/imagedraw_ellipse_RGB.png differ
diff --git a/Tests/images/imagedraw_ellipse_edge.png b/Tests/images/imagedraw_ellipse_edge.png
index 25a95a601..a2235115a 100644
Binary files a/Tests/images/imagedraw_ellipse_edge.png and b/Tests/images/imagedraw_ellipse_edge.png differ
diff --git a/Tests/images/imagedraw_ellipse_width.png b/Tests/images/imagedraw_ellipse_width.png
index ec0ca6731..54cfc291c 100644
Binary files a/Tests/images/imagedraw_ellipse_width.png and b/Tests/images/imagedraw_ellipse_width.png differ
diff --git a/Tests/images/imagedraw_ellipse_width_fill.png b/Tests/images/imagedraw_ellipse_width_fill.png
index 9b7be6029..4a1edc379 100644
Binary files a/Tests/images/imagedraw_ellipse_width_fill.png and b/Tests/images/imagedraw_ellipse_width_fill.png differ
diff --git a/Tests/images/imagedraw_ellipse_width_large.png b/Tests/images/imagedraw_ellipse_width_large.png
index 9d3c3326b..e95186019 100644
Binary files a/Tests/images/imagedraw_ellipse_width_large.png and b/Tests/images/imagedraw_ellipse_width_large.png differ
diff --git a/Tests/images/imagedraw_ellipse_zero_width.png b/Tests/images/imagedraw_ellipse_zero_width.png
index f14a279f2..6661b7d99 100644
Binary files a/Tests/images/imagedraw_ellipse_zero_width.png and b/Tests/images/imagedraw_ellipse_zero_width.png differ
diff --git a/Tests/test_imagedraw.py b/Tests/test_imagedraw.py
index f6eabb21a..cc29c097e 100644
--- a/Tests/test_imagedraw.py
+++ b/Tests/test_imagedraw.py
@@ -286,14 +286,14 @@ def test_ellipse_edge():
     draw = ImageDraw.Draw(im)
 
     # Act
-    draw.ellipse(((0, 0), (W - 1, H)), fill="white")
+    draw.ellipse(((0, 0), (W - 1, H - 1)), fill="white")
 
     # Assert
     assert_image_similar(im, Image.open("Tests/images/imagedraw_ellipse_edge.png"), 1)
 
 
 def test_ellipse_symmetric():
-    for bbox in [(25, 25, 76, 76), (25, 25, 75, 75)]:
+    for bbox in [(24, 25, 76, 75), (25, 25, 75, 75), (25, 24, 75, 76)]:
         im = Image.new("RGB", (101, 101))
         draw = ImageDraw.Draw(im)
         draw.ellipse(bbox, fill="green", outline="blue")
diff --git a/Tests/test_imagedraw2.py b/Tests/test_imagedraw2.py
index 72cbb79b8..c26cc0d94 100644
--- a/Tests/test_imagedraw2.py
+++ b/Tests/test_imagedraw2.py
@@ -79,7 +79,7 @@ def test_ellipse_edge():
     brush = ImageDraw2.Brush("white")
 
     # Act
-    draw.ellipse(((0, 0), (W - 1, H)), brush)
+    draw.ellipse(((0, 0), (W - 1, H - 1)), brush)
 
     # Assert
     assert_image_similar(im, Image.open("Tests/images/imagedraw_ellipse_edge.png"), 1)
diff --git a/src/libImaging/Draw.c b/src/libImaging/Draw.c
index ee3a3938e..34d1d4bc5 100644
--- a/src/libImaging/Draw.c
+++ b/src/libImaging/Draw.c
@@ -1047,78 +1047,76 @@ int8_t quarter_next(quarter_state* s, int32_t* ret_x, int32_t* ret_y) {
 typedef struct {
     quarter_state st_o, st_i;
     int32_t py, pl, pr;
-    int32_t cy, cl, cr;
-    // 0 - ready to take next quarter piece, 1, 2, 3 - returned 1, 2, 3 hlines
-    int8_t state;
+    int32_t cy[4], cl[4], cr[4];
+    int8_t bufcnt;
     int8_t finished;
+    int8_t leftmost;
 } ellipse_state;
 
 void ellipse_init(ellipse_state* s, int32_t a, int32_t b, int32_t w) {
-    s->state = 0;
+    s->bufcnt = 0;
+    s->leftmost = a % 2;
     quarter_init(&s->st_o, a, b);
     if (w < 1 || quarter_next(&s->st_o, &s->pr, &s->py) == -1) {
         s->finished = 1;
     } else {
         s->finished = 0;
         quarter_init(&s->st_i, a - 2 * (w - 1), b - 2 * (w - 1));
-        s->pl = a % 2;
+        s->pl = s->leftmost;
     }
 }
 
 int8_t ellipse_next(ellipse_state* s, int32_t* ret_x0, int32_t* ret_y, int32_t* ret_x1) {
-    switch (s->state) {
-        case 0: {
-            if (s->finished) {
-                return -1;
-            }
-            s->cy = s->py;
-            s->cl = s->pl;
-            s->cr = s->pr;
-            int32_t cx = 0, cy = 0;
-            int8_t next_ret;
-            while ((next_ret = quarter_next(&s->st_o, &cx, &cy)) != -1 && cy <= s->cy) {
-            }
-            s->pr = cx;
-            s->py = cy;
-            if (next_ret == -1) {
-                s->finished = 1;
-            }
-            while ((next_ret = quarter_next(&s->st_i, &cx, &cy)) != -1 && cy <= s->cy) {
-                s->cl = cx;
-            }
-            s->pl = next_ret == -1 ? 0 : cx;
-            *ret_x0 = -s->cr;
-            *ret_y = -s->cy;
-            *ret_x1 = -s->cl;
-            s->state = 1;
-            return 0;
-        }
-        case 1: {
-            *ret_x0 = s->cl;
-            *ret_y = -s->cy;
-            *ret_x1 = s->cr;
-            s->state = 2;
-            return 0;
-        }
-        case 2: {
-            *ret_x0 = -s->cr;
-            *ret_y = s->cy;
-            *ret_x1 = -s->cl;
-            s->state = 3;
-            return 0;
-        }
-        case 3: {
-            *ret_x0 = s->cl;
-            *ret_y = s->cy;
-            *ret_x1 = s->cr;
-            s->state = 0;
-            return 0;
-        }
-        default: {
-            assert(0);
+    if (s->bufcnt == 0) {
+        if (s->finished) {
             return -1;
         }
+        int32_t y = s->py;
+        int32_t l = s->pl;
+        int32_t r = s->pr;
+        int32_t cx = 0, cy = 0;
+        int8_t next_ret;
+        while ((next_ret = quarter_next(&s->st_o, &cx, &cy)) != -1 && cy <= y) {
+        }
+        if (next_ret == -1) {
+            s->finished = 1;
+        } else {
+            s->pr = cx;
+            s->py = cy;
+        }
+        while ((next_ret = quarter_next(&s->st_i, &cx, &cy)) != -1 && cy <= y) {
+            l = cx;
+        }
+        s->pl = next_ret == -1 ? s->leftmost : cx;
+
+        if ((l > 0 || l < r) && y > 0) {
+            s->cl[s->bufcnt] = l == 0 ? 2 : l;
+            s->cy[s->bufcnt] = y;
+            s->cr[s->bufcnt] = r;
+            ++s->bufcnt;
+        }
+        if (y > 0) {
+            s->cl[s->bufcnt] = -r;
+            s->cy[s->bufcnt] = y;
+            s->cr[s->bufcnt] = -l;
+            ++s->bufcnt;
+        }
+        if (l > 0 || l < r) {
+            s->cl[s->bufcnt] = l == 0 ? 2 : l;
+            s->cy[s->bufcnt] = -y;
+            s->cr[s->bufcnt] = r;
+            ++s->bufcnt;
+        }
+        s->cl[s->bufcnt] = -r;
+        s->cy[s->bufcnt] = -y;
+        s->cr[s->bufcnt] = -l;
+        ++s->bufcnt;
     }
+    --s->bufcnt;
+    *ret_x0 = s->cl[s->bufcnt];
+    *ret_y = s->cy[s->bufcnt];
+    *ret_x1 = s->cr[s->bufcnt];
+    return 0;
 }
 
 static int