diff --git a/src/d3d/shaders/matfx_env_PS.h b/src/d3d/shaders/matfx_env_PS.h
index 41d5678..2224b68 100644
--- a/src/d3d/shaders/matfx_env_PS.h
+++ b/src/d3d/shaders/matfx_env_PS.h
@@ -32,13 +32,14 @@
     max r1.xyz, v0, c2
     mul r1.xyz, r1, c1.x
     mul r0.xyz, r0, r1
+    mul r0.xyz, r0, t0.z
     add r1.xyz, v0, -c0
     mad r1.xyz, t0.z, r1, c0
     mad r0.xyz, r1, v0.w, r0
     mov r0.w, v0.w
     mov oC0, r0
 
-// approximately 9 instruction slots used (1 texture, 8 arithmetic)
+// approximately 10 instruction slots used (1 texture, 9 arithmetic)
 #endif
 
 const BYTE g_ps20_main[] =
@@ -105,19 +106,22 @@ const BYTE g_ps20_main[] =
     228, 128,   1,   0,   0, 160, 
       5,   0,   0,   3,   0,   0, 
       7, 128,   0,   0, 228, 128, 
-      1,   0, 228, 128,   2,   0, 
-      0,   3,   1,   0,   7, 128, 
-      0,   0, 228, 144,   0,   0, 
-    228, 161,   4,   0,   0,   4, 
+      1,   0, 228, 128,   5,   0, 
+      0,   3,   0,   0,   7, 128, 
+      0,   0, 228, 128,   0,   0, 
+    170, 176,   2,   0,   0,   3, 
       1,   0,   7, 128,   0,   0, 
-    170, 176,   1,   0, 228, 128, 
-      0,   0, 228, 160,   4,   0, 
-      0,   4,   0,   0,   7, 128, 
+    228, 144,   0,   0, 228, 161, 
+      4,   0,   0,   4,   1,   0, 
+      7, 128,   0,   0, 170, 176, 
       1,   0, 228, 128,   0,   0, 
-    255, 144,   0,   0, 228, 128, 
-      1,   0,   0,   2,   0,   0, 
-      8, 128,   0,   0, 255, 144, 
-      1,   0,   0,   2,   0,   8, 
-     15, 128,   0,   0, 228, 128, 
-    255, 255,   0,   0
+    228, 160,   4,   0,   0,   4, 
+      0,   0,   7, 128,   1,   0, 
+    228, 128,   0,   0, 255, 144, 
+      0,   0, 228, 128,   1,   0, 
+      0,   2,   0,   0,   8, 128, 
+      0,   0, 255, 144,   1,   0, 
+      0,   2,   0,   8,  15, 128, 
+      0,   0, 228, 128, 255, 255, 
+      0,   0
 };
diff --git a/src/d3d/shaders/matfx_env_PS.hlsl b/src/d3d/shaders/matfx_env_PS.hlsl
index 3508351..767e313 100644
--- a/src/d3d/shaders/matfx_env_PS.hlsl
+++ b/src/d3d/shaders/matfx_env_PS.hlsl
@@ -16,14 +16,15 @@ float4 colorClamp : register(c2);
 float4 main(VS_out input) : COLOR
 {
 	float4 pass1 = input.Color;
+	float4 envColor = max(pass1, colorClamp);
 #ifdef TEX
 	pass1 *= tex2D(diffTex, input.TexCoord0.xy);
 #endif
 
-	float4 envColor = max(pass1, colorClamp);
 	float4 pass2 = envColor*shininess*tex2D(envTex, input.TexCoord1.xy);
 
 	pass1.rgb = lerp(fogColor.rgb, pass1.rgb, input.TexCoord0.z);
+	pass2.rgb = lerp(float3(0.0, 0.0, 0.0), pass2.rgb, input.TexCoord0.z);
 
 	// We simulate drawing this in two passes.
 	// First pass with standard blending, second with addition
diff --git a/src/d3d/shaders/matfx_env_tex_PS.h b/src/d3d/shaders/matfx_env_tex_PS.h
index 4559954..f6054a9 100644
--- a/src/d3d/shaders/matfx_env_tex_PS.h
+++ b/src/d3d/shaders/matfx_env_tex_PS.h
@@ -31,18 +31,19 @@
     dcl v0
     dcl_2d s0
     dcl_2d s1
-    texld r0, t0, s0
-    texld r1, t1, s1
-    mad r2.xyz, v0, r0, -c0
-    mul r0, r0, v0
-    mad r2.xyz, t0.z, r2, c0
-    max r3.xyz, r0, c2
-    mul r3.xyz, r3, c1.x
-    mul r1.xyz, r1, r3
-    mad r0.xyz, r2, r0.w, r1
-    mov oC0, r0
+    texld r0, t1, s1
+    texld r1, t0, s0
+    max r2.xyz, v0, c2
+    mul r2.xyz, r2, c1.x
+    mul r0.xyz, r0, r2
+    mul r0.xyz, r0, t0.z
+    mad r1.xyz, v0, r1, -c0
+    mul r2.w, r1.w, v0.w
+    mad r1.xyz, t0.z, r1, c0
+    mad r2.xyz, r1, r2.w, r0
+    mov oC0, r2
 
-// approximately 10 instruction slots used (2 texture, 8 arithmetic)
+// approximately 11 instruction slots used (2 texture, 9 arithmetic)
 #endif
 
 const BYTE g_ps20_main[] =
@@ -109,32 +110,35 @@ const BYTE g_ps20_main[] =
      15, 160,  31,   0,   0,   2, 
       0,   0,   0, 144,   1,   8, 
      15, 160,  66,   0,   0,   3, 
-      0,   0,  15, 128,   0,   0, 
-    228, 176,   0,   8, 228, 160, 
+      0,   0,  15, 128,   1,   0, 
+    228, 176,   1,   8, 228, 160, 
      66,   0,   0,   3,   1,   0, 
-     15, 128,   1,   0, 228, 176, 
-      1,   8, 228, 160,   4,   0, 
-      0,   4,   2,   0,   7, 128, 
-      0,   0, 228, 144,   0,   0, 
-    228, 128,   0,   0, 228, 161, 
+     15, 128,   0,   0, 228, 176, 
+      0,   8, 228, 160,  11,   0, 
+      0,   3,   2,   0,   7, 128, 
+      0,   0, 228, 144,   2,   0, 
+    228, 160,   5,   0,   0,   3, 
+      2,   0,   7, 128,   2,   0, 
+    228, 128,   1,   0,   0, 160, 
       5,   0,   0,   3,   0,   0, 
-     15, 128,   0,   0, 228, 128, 
-      0,   0, 228, 144,   4,   0, 
-      0,   4,   2,   0,   7, 128, 
-      0,   0, 170, 176,   2,   0, 
-    228, 128,   0,   0, 228, 160, 
-     11,   0,   0,   3,   3,   0, 
       7, 128,   0,   0, 228, 128, 
-      2,   0, 228, 160,   5,   0, 
-      0,   3,   3,   0,   7, 128, 
-      3,   0, 228, 128,   1,   0, 
-      0, 160,   5,   0,   0,   3, 
-      1,   0,   7, 128,   1,   0, 
-    228, 128,   3,   0, 228, 128, 
-      4,   0,   0,   4,   0,   0, 
-      7, 128,   2,   0, 228, 128, 
-      0,   0, 255, 128,   1,   0, 
-    228, 128,   1,   0,   0,   2, 
-      0,   8,  15, 128,   0,   0, 
-    228, 128, 255, 255,   0,   0
+      2,   0, 228, 128,   5,   0, 
+      0,   3,   0,   0,   7, 128, 
+      0,   0, 228, 128,   0,   0, 
+    170, 176,   4,   0,   0,   4, 
+      1,   0,   7, 128,   0,   0, 
+    228, 144,   1,   0, 228, 128, 
+      0,   0, 228, 161,   5,   0, 
+      0,   3,   2,   0,   8, 128, 
+      1,   0, 255, 128,   0,   0, 
+    255, 144,   4,   0,   0,   4, 
+      1,   0,   7, 128,   0,   0, 
+    170, 176,   1,   0, 228, 128, 
+      0,   0, 228, 160,   4,   0, 
+      0,   4,   2,   0,   7, 128, 
+      1,   0, 228, 128,   2,   0, 
+    255, 128,   0,   0, 228, 128, 
+      1,   0,   0,   2,   0,   8, 
+     15, 128,   2,   0, 228, 128, 
+    255, 255,   0,   0
 };
diff --git a/src/d3d/shaders/skin_VS.hlsl b/src/d3d/shaders/skin_VS.hlsl
new file mode 100644
index 0000000..7e19c08
--- /dev/null
+++ b/src/d3d/shaders/skin_VS.hlsl
@@ -0,0 +1,66 @@
+#include "standardConstants.h"
+
+float4x4 boneMatrices[64] : register(c41);
+
+struct VS_in
+{
+	float4 Position		: POSITION;
+	float3 Normal		: NORMAL;
+	float2 TexCoord		: TEXCOORD0;
+	float4 Prelight		: COLOR0;
+	float4 Weights		: BLENDWEIGHT;
+	int4 Indices		: BLENDINDICES;
+};
+
+struct VS_out {
+	float4 Position		: POSITION;
+	float3 TexCoord0	: TEXCOORD0;	// also fog
+	float4 Color		: COLOR0;
+};
+
+
+VS_out main(in VS_in input)
+{
+	VS_out output;
+
+	int j;
+	float3 SkinVertex = float3(0.0, 0.0, 0.0);
+	float3 SkinNormal = float3(0.0, 0.0, 0.0);
+	for(j = 0; j < 4; j++){
+		SkinVertex += mul(boneMatrices[input.Indices[j]], input.Position).xyz * input.Weights[j];
+		SkinNormal += mul((float3x3)boneMatrices[input.Indices[j]], input.Normal).xyz * input.Weights[j];
+	}
+
+	output.Position = mul(combinedMat, float4(SkinVertex, 1.0));
+	float3 Vertex = mul(worldMat, float4(SkinVertex, 1.0)).xyz;
+	float3 Normal = mul(normalMat, SkinNormal);
+
+	output.TexCoord0.xy = input.TexCoord;
+
+	output.Color = float4(0.0, 0.0, 0.0, 1.0);
+	if(surfPrelight > 0.0)
+		output.Color = input.Prelight;
+
+	output.Color.rgb += ambientLight.rgb * surfAmbient;
+
+	int i;
+#ifdef DIRECTIONALS
+	for(i = 0; i < numDirLights; i++)
+		output.Color.xyz += DoDirLight(lights[i+firstDirLight], Normal)*surfDiffuse;
+#endif
+#ifdef POINTLIGHTS
+	for(i = 0; i < numPointLights; i++)
+		output.Color.xyz += DoPointLight(lights[i+firstPointLight], Vertex.xyz, Normal)*surfDiffuse;
+#endif
+#ifdef SPOTLIGHTS
+	for(i = 0; i < numSpotLights; i++)
+		output.Color.xyz += DoSpotLight(lights[i+firstSpotLight], Vertex.xyz, Normal)*surfDiffuse;
+#endif
+	// PS2 clamps before material color
+	output.Color = clamp(output.Color, 0.0, 1.0);
+	output.Color *= matCol;
+
+	output.TexCoord0.z = clamp((output.Position.w - fogEnd)*fogRange, fogDisable, 1.0);
+
+	return output;
+}