diff --git a/src/nvimage/DirectDrawSurface.cpp b/src/nvimage/DirectDrawSurface.cpp
index 67840f3..41b4f0b 100644
--- a/src/nvimage/DirectDrawSurface.cpp
+++ b/src/nvimage/DirectDrawSurface.cpp
@@ -639,15 +639,7 @@ void DDSHeader::setFourCC(uint8 c0, uint8 c1, uint8 c2, uint8 c3)
 	this->pf.flags = DDPF_FOURCC;
 	this->pf.fourcc = MAKEFOURCC(c0, c1, c2, c3);
 	
-	if (this->pf.fourcc == FOURCC_ATI2)
-	{
-		this->pf.bitcount = FOURCC_A2XY;
-	}
-	else
-	{
-		this->pf.bitcount = 0;
-	}
-	
+	this->pf.bitcount = 0;
 	this->pf.rmask = 0;
 	this->pf.gmask = 0;
 	this->pf.bmask = 0;
@@ -667,6 +659,11 @@ void DDSHeader::setFormatCode(uint32 code)
 	this->pf.amask = 0;
 }
 
+void DDSHeader::setSwizzleCode(uint8 c0, uint8 c1, uint8 c2, uint8 c3)
+{
+	this->pf.bitcount = MAKEFOURCC(c0, c1, c2, c3);
+}
+
 
 void DDSHeader::setPixelFormat(uint bitcount, uint rmask, uint gmask, uint bmask, uint amask)
 {
diff --git a/src/nvimage/DirectDrawSurface.h b/src/nvimage/DirectDrawSurface.h
index c5fea9c..918a6f9 100644
--- a/src/nvimage/DirectDrawSurface.h
+++ b/src/nvimage/DirectDrawSurface.h
@@ -94,6 +94,7 @@ namespace nv
 		void setPitch(uint pitch);
 		void setFourCC(uint8 c0, uint8 c1, uint8 c2, uint8 c3);
 		void setFormatCode(uint code);
+		void setSwizzleCode(uint8 c0, uint8 c1, uint8 c2, uint8 c3);
 		void setPixelFormat(uint bitcount, uint rmask, uint gmask, uint bmask, uint amask);
 		void setDX10Format(uint format);
 		void setNormalFlag(bool b);
diff --git a/src/nvtt/CompressDXT.cpp b/src/nvtt/CompressDXT.cpp
index 19124dc..fe83607 100644
--- a/src/nvtt/CompressDXT.cpp
+++ b/src/nvtt/CompressDXT.cpp
@@ -371,6 +371,9 @@ void nv::SlowCompressor::compressDXT5n(const CompressionOptions::Private & compr
 	ColorBlock rgba;
 	BlockDXT5 block;
 	
+	squish::WeightedClusterFit fit;
+	fit.SetMetric(0, 1, 0);
+
 	for (uint y = 0; y < h; y += 4) {
 		for (uint x = 0; x < w; x += 4) {
 			
@@ -389,7 +392,18 @@ void nv::SlowCompressor::compressDXT5n(const CompressionOptions::Private & compr
 			}
 			
 			// Compress Y.
-			OptimalCompress::compressDXT1G(rgba, &block.color);
+			//OptimalCompress::compressDXT1G(rgba, &block.color);
+
+			/*if (rgba.isSingleColor())
+			{
+				OptimalCompress::compressDXT1G(rgba.color(0), &block.color);
+			}
+			else*/
+			{
+				squish::ColourSet colours((uint8 *)rgba.colors(), squish::kWeightColourByAlpha);
+				fit.SetColourSet(&colours, 0);
+				fit.Compress(&block.color);
+			}
 			
 			if (outputOptions.outputHandler != NULL) {
 				outputOptions.outputHandler->writeData(&block, sizeof(block));
diff --git a/src/nvtt/Compressor.cpp b/src/nvtt/Compressor.cpp
index 4ff6840..31b7423 100644
--- a/src/nvtt/Compressor.cpp
+++ b/src/nvtt/Compressor.cpp
@@ -453,14 +453,21 @@ bool Compressor::Private::outputHeader(const InputOptions::Private & inputOption
 				}
 				else if (compressionOptions.format == Format_DXT5n) {
 					header.setFourCC('D', 'X', 'T', '5');
-					if (inputOptions.isNormalMap) header.setNormalFlag(true);
+					if (inputOptions.isNormalMap) {
+						header.setNormalFlag(true);
+						header.setSwizzleCode('A', '2', 'D', '5');
+						//header.setSwizzleCode('x', 'G', 'x', 'R');
+					}
 				}
 				else if (compressionOptions.format == Format_BC4) {
 					header.setFourCC('A', 'T', 'I', '1');
 				}
 				else if (compressionOptions.format == Format_BC5) {
 					header.setFourCC('A', 'T', 'I', '2');
-					if (inputOptions.isNormalMap) header.setNormalFlag(true);
+					if (inputOptions.isNormalMap) {
+						header.setNormalFlag(true);
+						header.setSwizzleCode('A', '2', 'X', 'Y');
+					}
 				}
 				else if (compressionOptions.format == Format_CTX1) {
 					header.setFourCC('C', 'T', 'X', '1');
@@ -905,7 +912,8 @@ bool Compressor::Private::compressMipmap(const Mipmap & mipmap, const InputOptio
 				{
 					nvDebugCheck(cudaSupported);
 					cuda->setImage(image, inputOptions.alphaMode);
-					cuda->compressDXT1(compressionOptions, outputOptions);
+					//cuda->compressDXT1(compressionOptions, outputOptions);
+					cuda->compressDXT1_Tex(compressionOptions, outputOptions);
 				}
 				else
 				{
@@ -993,7 +1001,16 @@ bool Compressor::Private::compressMipmap(const Mipmap & mipmap, const InputOptio
 			}
 			else
 			{
-				slow.compressDXT5n(compressionOptions, outputOptions);
+				/*if (cudaEnabled)
+				{
+					nvDebugCheck(cudaSupported);
+					cuda->setImage(image, inputOptions.alphaMode);
+					cuda->compressDXT5n(compressionOptions, outputOptions);
+				}
+				else*/
+				{
+					slow.compressDXT5n(compressionOptions, outputOptions);
+				}
 			}
 		}
 		else if (compressionOptions.format == Format_BC4)