diff --git a/libswscale/input.c b/libswscale/input.c
index e636eac1f6..b4855093a0 100644
--- a/libswscale/input.c
+++ b/libswscale/input.c
@@ -309,7 +309,8 @@ static void monowhite2Y_c(uint8_t *dst, const uint8_t *src,
                           int width, uint32_t *unused)
 {
     int i, j;
-    for (i=0; i<width/8; i++) {
+    width = (width + 7) >> 3;
+    for (i = 0; i < width; i++) {
         int d= ~src[i];
         for(j=0; j<8; j++)
             dst[8*i+j]= ((d>>(7-j))&1)*255;
@@ -320,7 +321,8 @@ static void monoblack2Y_c(uint8_t *dst, const uint8_t *src,
                           int width, uint32_t *unused)
 {
     int i, j;
-    for (i=0; i<width/8; i++) {
+    width = (width + 7) >> 3;
+    for (i = 0; i < width; i++) {
         int d= src[i];
         for(j=0; j<8; j++)
             dst[8*i+j]= ((d>>(7-j))&1)*255;
diff --git a/libswscale/output.c b/libswscale/output.c
index 533fcd945d..f1acf9ec92 100644
--- a/libswscale/output.c
+++ b/libswscale/output.c
@@ -317,7 +317,7 @@ yuv2mono_X_c_template(SwsContext *c, const int16_t *lumFilter,
     int i;
     unsigned acc = 0;
 
-    for (i = 0; i < dstW - 1; i += 2) {
+    for (i = 0; i < dstW; i += 2) {
         int j;
         int Y1 = 1 << 18;
         int Y2 = 1 << 18;
@@ -338,6 +338,10 @@ yuv2mono_X_c_template(SwsContext *c, const int16_t *lumFilter,
             output_pixel(*dest++, acc);
         }
     }
+
+    if (i & 6) {
+        output_pixel(*dest, acc);
+    }
 }
 
 static av_always_inline void
@@ -352,7 +356,7 @@ yuv2mono_2_c_template(SwsContext *c, const int16_t *buf[2],
     int  yalpha1 = 4095 - yalpha;
     int i;
 
-    for (i = 0; i < dstW - 7; i += 8) {
+    for (i = 0; i < dstW; i += 8) {
         int Y, acc = 0;
 
         Y = (buf0[i + 0] * yalpha1 + buf1[i + 0] * yalpha) >> 19;
@@ -385,7 +389,7 @@ yuv2mono_1_c_template(SwsContext *c, const int16_t *buf0,
     const uint8_t * const d128 = dither_8x8_220[y & 7];
     int i;
 
-    for (i = 0; i < dstW - 7; i += 8) {
+    for (i = 0; i < dstW; i += 8) {
         int acc = 0;
 
         accumulate_bit(acc, (buf0[i + 0] >> 7) + d128[0]);