diff --git a/src/capture_line_ntsc_8bpp.S b/src/capture_line_ntsc_8bpp.S
index 707e45bc..4ade397b 100644
--- a/src/capture_line_ntsc_8bpp.S
+++ b/src/capture_line_ntsc_8bpp.S
@@ -1320,56 +1320,40 @@ loop_8bppd_auto:
 .macro CAPTURE_SIX_BITS_16BPP_0 reg
         // Pixel 0 in GPIO  7.. 2 ->  7.. 0
         // Pixel 1 in GPIO 13.. 8 -> 15.. 8
-        mov    \reg, #0
-        tst    r8, #(0x01 << PIXEL_BASE)
-        orrne  \reg, #0x04
-        tst    r8, #(0x02 << PIXEL_BASE)
-        orrne  \reg, #0x02
-        tst    r8, #(0x04 << PIXEL_BASE)
-        orrne  \reg, #0x01
+        and    r9, r8, #(0x07 << PIXEL_BASE)
         tst    r8, #(0x10 << PIXEL_BASE)
-        orrne  \reg, #0x08
+        orrne  r9, r9, #(0x08 << PIXEL_BASE)
+        mov    \reg, r9, lsr #PIXEL_BASE
 .endm
 
 .macro CAPTURE_SIX_BITS_16BPP_1 reg
         // Pixel 2 in GPIO  7.. 2 -> 23..16
         // Pixel 3 in GPIO 13.. 8 -> 31..24
-        tst    r8, #(0x01 << (PIXEL_BASE + 6))
-        orrne  \reg, #0x0400
-        tst    r8, #(0x02 << (PIXEL_BASE + 6))
-        orrne  \reg, #0x0200
-        tst    r8, #(0x04 << (PIXEL_BASE + 6))
-        orrne  \reg, #0x0100
+        and    r9, r8, #(0x07 << (PIXEL_BASE + 6))
         tst    r8, #(0x10 << (PIXEL_BASE + 6))
-        orrne  \reg, #0x0800
+        orrne  r9, r9, #(0x08 << (PIXEL_BASE + 6))
+        orr    \reg, \reg, r9, lsl #(8 - (PIXEL_BASE + 6))
 .endm
 
 .macro CAPTURE_SIX_BITS_16BPP_2 reg
         // Pixel 0 in GPIO  7.. 2 ->  7.. 0
         // Pixel 1 in GPIO 13.. 8 -> 15.. 8
-        tst    r8, #(0x01 << PIXEL_BASE)
-        orrne  \reg, #0x040000
-        tst    r8, #(0x02 << PIXEL_BASE)
-        orrne  \reg, #0x020000
-        tst    r8, #(0x04 << PIXEL_BASE)
-        orrne  \reg, #0x010000
+        and    r9, r8, #(0x07 << PIXEL_BASE)
         tst    r8, #(0x10 << PIXEL_BASE)
-        orrne  \reg, #0x080000
+        orrne  r9, r9, #(0x08 << PIXEL_BASE)
+        orr    \reg, \reg, r9, lsl #(16 - PIXEL_BASE)
 .endm
 
 .macro CAPTURE_SIX_BITS_16BPP_3 reg
         // Pixel 2 in GPIO  7.. 2 -> 23..16
         // Pixel 3 in GPIO 13.. 8 -> 31..24
-        tst    r8, #(0x01 << (PIXEL_BASE + 6))
-        orrne  \reg, #0x04000000
-        tst    r8, #(0x02 << (PIXEL_BASE + 6))
-        orrne  \reg, #0x02000000
-        tst    r8, #(0x04 << (PIXEL_BASE + 6))
-        orrne  \reg, #0x01000000
+        and    r9, r8, #(0x07 << (PIXEL_BASE + 6))
         tst    r8, #(0x10 << (PIXEL_BASE + 6))
-        orrne  \reg, #0x08000000
+        orrne  r9, r9, #(0x08 << (PIXEL_BASE + 6))
+        orr    \reg, \reg, r9, lsl #(24 - (PIXEL_BASE + 6))
 .endm
 
+
 .global cga_process_artifact
 .global cga_render_words
 
diff --git a/src/vid_cga_comp.c b/src/vid_cga_comp.c
index 7fce5035..43795335 100644
--- a/src/vid_cga_comp.c
+++ b/src/vid_cga_comp.c
@@ -76,6 +76,13 @@ double max_v;
 int video_ri, video_rq, video_gi, video_gq, video_bi, video_bq;
 static int ntsc_type;
 
+unsigned int swap_R_B(int IRGB) {
+    //r & b are swapped in RGBtoHDMI so rearrange to match
+    int r_new = (IRGB & 0x04) >> 2;
+    int ig_new = (IRGB & 0x0a);
+    int b_new = (IRGB & 0x01) << 2;
+    return b_new | ig_new | r_new;
+}
 
 void update_cga16_color() {
     int x;
@@ -155,8 +162,8 @@ void update_cga16_color() {
 
         for (x = 0; x < 1024; ++x) {
                 int phase = (x + ntsc_pixel_phase) & 3;
-                int right = (x >> 2) & 15;
-                int left = (x >> 6) & 15;
+                int right = swap_R_B((x >> 2) & 0x0f);
+                int left = swap_R_B((x >> 6) & 0x0f);
                 int rc = right;
                 int lc = left;
                 //if ((cgamode & 4) != 0) {
@@ -176,8 +183,8 @@ void update_cga16_color() {
                 CGA_Composite_Table[x] =  (int) (v*mode_contrast + mode_brightness);
         }
 
-        i = CGA_Composite_Table[6*68] - CGA_Composite_Table[6*68 + 2];
-        q = CGA_Composite_Table[6*68 + 1] - CGA_Composite_Table[6*68 + 3];
+        i = CGA_Composite_Table[3*68] - CGA_Composite_Table[3*68 + 2];
+        q = CGA_Composite_Table[3*68 + 1] - CGA_Composite_Table[3*68 + 3];
 
         a = tau*(33 + 90 + hue_offset + mode_hue)/360.0;
         c = cos(a);