diff --git a/desmume/src/commandline.cpp b/desmume/src/commandline.cpp
index 851d88558..2a072c9e3 100644
--- a/desmume/src/commandline.cpp
+++ b/desmume/src/commandline.cpp
@@ -28,7 +28,7 @@
 #include "NDSSystem.h"
 #include "utils/xstring.h"
 
-int scanline_filter_a = 2, scanline_filter_b = 4;
+int scanline_filter_a = 0, scanline_filter_b = 2, scanline_filter_c = 2, scanline_filter_d = 4;
 int _commandline_linux_nojoy = 0;
 
 CommandLine::CommandLine()
@@ -91,8 +91,10 @@ void CommandLine::loadCommonOptions()
 		{ "bios-swi", 0, 0, G_OPTION_ARG_INT, &_bios_swi, "Uses SWI from the provided bios files", "BIOS_SWI"},
 		{ "spu-advanced", 0, 0, G_OPTION_ARG_INT, &_spu_advanced, "Uses advanced SPU capture functions", "SPU_ADVANCED"},
 		{ "num-cores", 0, 0, G_OPTION_ARG_INT, &_num_cores, "Override numcores detection and use this many", "NUM_CORES"},
-		{ "scanline-filter-a", 0, 0, G_OPTION_ARG_INT, &scanline_filter_a, "Intensity of fadeout for scanlines filter (edge) (default 2)", "SCANLINE_FILTER_A"},
-		{ "scanline-filter-b", 0, 0, G_OPTION_ARG_INT, &scanline_filter_b, "Intensity of fadeout for scanlines filter (corner) (default 4)", "SCANLINE_FILTER_B"},
+		{ "scanline-filter-a", 0, 0, G_OPTION_ARG_INT, &scanline_filter_a, "Intensity of fadeout for scanlines filter (topleft) (default 0)", "SCANLINE_FILTER_A"},
+		{ "scanline-filter-b", 0, 0, G_OPTION_ARG_INT, &scanline_filter_b, "Intensity of fadeout for scanlines filter (topright) (default 2)", "SCANLINE_FILTER_B"},
+		{ "scanline-filter-c", 0, 0, G_OPTION_ARG_INT, &scanline_filter_c, "Intensity of fadeout for scanlines filter (bottomleft) (default 2)", "SCANLINE_FILTER_C"},
+		{ "scanline-filter-d", 0, 0, G_OPTION_ARG_INT, &scanline_filter_d, "Intensity of fadeout for scanlines filter (bottomright) (default 4)", "SCANLINE_FILTER_D"},
 		{ "rigorous-timing", 0, 0, G_OPTION_ARG_INT, &_rigorous_timing, "Use some rigorous timings instead of unrealistically generous (default 0)", "RIGOROUS_TIMING"},
 		{ "advanced-timing", 0, 0, G_OPTION_ARG_INT, &_advanced_timing, "Use advanced BUS-level timing (default 1)", "ADVANCED_TIMING"},
 		{ "slot1", 0, 0, G_OPTION_ARG_STRING, &_slot1, "Device to load in slot 1 (default retail)", "SLOT1"},
diff --git a/desmume/src/filter/scanline.cpp b/desmume/src/filter/scanline.cpp
index b14055f37..6a034b792 100644
--- a/desmume/src/filter/scanline.cpp
+++ b/desmume/src/filter/scanline.cpp
@@ -24,57 +24,29 @@
 
 typedef u64 uint64;
 
-extern CACHE_ALIGN u16 fadeOutColors[17][0x8000];
+extern int scanline_filter_a, scanline_filter_b, scanline_filter_c, scanline_filter_d;
+static int fac_a, fac_b, fac_c, fac_d;
 
-extern int scanline_filter_a, scanline_filter_b;
-static int fac_a, fac_b;
-
-FORCEINLINE void ScanLine16( uint16 *lpDst, uint16 *lpSrc, unsigned int Width){
-	while(Width--){
-		*lpDst++ = *lpSrc;
-		*lpDst++ = fadeOutColors[scanline_filter_a][(*lpSrc++)];
-	}
-}
-
-FORCEINLINE void ScanLine16_2( uint16 *lpDst, uint16 *lpSrc, unsigned int Width){
-	while(Width--){
-		*lpDst++ = fadeOutColors[scanline_filter_a][(*lpSrc)];
-		*lpDst++ = fadeOutColors[scanline_filter_b][(*lpSrc++)];
-	}
-}
-
-FORCEINLINE void ScanLine32( uint32 *lpDst, uint32 *lpSrc, unsigned int Width){
-	while(Width--){
-		*lpDst++ = *lpSrc;
+FORCEINLINE void ScanLine32( uint32 *lpDst, uint32 *lpSrc, unsigned int Width, int fac_left, int fac_right)
+{
+	while(Width--)
+	{
 		u8* u8dst = (u8*)lpDst;
 		u8* u8src = (u8*)lpSrc;
-		*u8dst++ = *u8src++ * fac_a / 16;
-		*u8dst++ = *u8src++ * fac_a / 16;
-		*u8dst++ = *u8src++ * fac_a / 16;
-		lpDst++; 
-		lpSrc++;
-	}
-}
-
-FORCEINLINE void ScanLine32_2( uint32 *lpDst, uint32 *lpSrc, unsigned int Width){
-	while(Width--){
-		u8* u8dst = (u8*)lpDst;
-		u8* u8src = (u8*)lpSrc;
-		*u8dst++ = *u8src++ * fac_a / 16;
-		*u8dst++ = *u8src++ * fac_a / 16;
-		*u8dst++ = *u8src++ * fac_a / 16;
+		*u8dst++ = *u8src++ * fac_left / 16;
+		*u8dst++ = *u8src++ * fac_left / 16;
+		*u8dst++ = *u8src++ * fac_left / 16;
 		u8dst++;
 		u8src = (u8*)lpSrc;
-		*u8dst++ = *u8src++ * fac_b / 16;
-		*u8dst++ = *u8src++ * fac_b / 16;
-		*u8dst++ = *u8src++ * fac_b / 16;
+		*u8dst++ = *u8src++ * fac_right / 16;
+		*u8dst++ = *u8src++ * fac_right / 16;
+		*u8dst++ = *u8src++ * fac_right / 16;
 		u8dst++; u8src++;
 		lpDst+=2; 
 		lpSrc++;
 	}
 }
 
-
 FORCEINLINE void DoubleLine32( uint32 *lpDst, uint32 *lpSrc, unsigned int Width){
 	while(Width--){
 		*lpDst++ = *lpSrc;
@@ -86,6 +58,8 @@ void RenderScanline( SSurface Src, SSurface Dst)
 {
 	fac_a = (16-scanline_filter_a);
 	fac_b = (16-scanline_filter_b);
+	fac_c = (16-scanline_filter_c);
+	fac_d = (16-scanline_filter_d);
 	unsigned int H;
 
 	const uint32 srcHeight = Src.Height;
@@ -96,9 +70,12 @@ void RenderScanline( SSurface Src, SSurface Dst)
 	const unsigned int dstPitch = Dst.Pitch >> 1;
 	u32 *lpDst = (u32*)Dst.Surface;
 	for (H = 0; H < srcHeight; H++, lpSrc += srcPitch)
-		ScanLine32 (lpDst, lpSrc, Src.Width), lpDst += dstPitch,
-		ScanLine32_2 (lpDst, lpSrc, Src.Width), lpDst += dstPitch;
-		//memset (lpDst, 0, 512*2), lpDst += dstPitch;
+	{
+		ScanLine32(lpDst, lpSrc, Src.Width, fac_a, fac_b);
+		lpDst += dstPitch;
+		ScanLine32(lpDst, lpSrc, Src.Width, fac_c, fac_d);
+		lpDst += dstPitch;
+	}
 }
 
 void RenderNearest2X (SSurface Src, SSurface Dst)