diff --git a/bsnes.exe b/bsnes.exe
index 69c84606..41af013f 100644
Binary files a/bsnes.exe and b/bsnes.exe differ
diff --git a/src/snes/ppu/sppu/screen/screen.cpp b/src/snes/ppu/sppu/screen/screen.cpp
index 697d8913..2bf0ed82 100644
--- a/src/snes/ppu/sppu/screen/screen.cpp
+++ b/src/snes/ppu/sppu/screen/screen.cpp
@@ -103,7 +103,7 @@ uint16 sPPU::Screen::get_pixel(bool swap) {
     source_sub = OAM;
   }
   if(priority_sub == 0) {
-    if(self.regs.bgmode == 5 || self.regs.bgmode == 6) {
+    if(self.regs.pseudo_hires == true || self.regs.bgmode == 5 || self.regs.bgmode == 6) {
       color_sub = get_color(0);
     } else {
       color_sub = (regs.color_b << 10) + (regs.color_g << 5) + (regs.color_r << 0);
diff --git a/src/snes/ppu/sppu/serialization.cpp b/src/snes/ppu/sppu/serialization.cpp
index 73efc62c..0a6a0a54 100644
--- a/src/snes/ppu/sppu/serialization.cpp
+++ b/src/snes/ppu/sppu/serialization.cpp
@@ -118,7 +118,6 @@ void sPPU::Sprite::serialize(serializer &s) {
   s.integer(t.x);
   s.integer(t.y);
 
-  s.integer(t.active_sprite);
   s.integer(t.item_count);
   s.integer(t.tile_count);
 
@@ -127,11 +126,13 @@ void sPPU::Sprite::serialize(serializer &s) {
     s.array(t.item[n]);
     for(unsigned i = 0; i < 34; i++) {
       s.integer(t.tile[n][i].x);
-      s.integer(t.tile[n][i].y);
       s.integer(t.tile[n][i].priority);
       s.integer(t.tile[n][i].palette);
-      s.integer(t.tile[n][i].tile);
       s.integer(t.tile[n][i].hflip);
+      s.integer(t.tile[n][i].d0);
+      s.integer(t.tile[n][i].d1);
+      s.integer(t.tile[n][i].d2);
+      s.integer(t.tile[n][i].d3);
     }
   }
 
diff --git a/src/snes/ppu/sppu/sppu.cpp b/src/snes/ppu/sppu/sppu.cpp
index dfcf0dd7..acb2bc4e 100644
--- a/src/snes/ppu/sppu/sppu.cpp
+++ b/src/snes/ppu/sppu/sppu.cpp
@@ -26,7 +26,7 @@ void sPPU::enter() {
     scanline();
     add_clocks(88);
 
-    if(vcounter() >= 1 && vcounter() <= (!regs.overscan ? 224 : 239)) {
+    if(vcounter() <= (!regs.overscan ? 224 : 239)) {
       for(unsigned n = 0; n < 256; n++) {
         bg1.run();
         bg2.run();
@@ -43,11 +43,14 @@ void sPPU::enter() {
         screen.run();
         add_clocks(2);
       }
+
+      add_clocks(22);
+      oam.tilefetch();
     } else {
-      add_clocks(1024);
+      add_clocks(1024 + 22 + 136);
     }
 
-    add_clocks(lineclocks() - 88 - 1024);
+    add_clocks(lineclocks() - 88 - 1024 - 22 - 136);
   }
 }
 
diff --git a/src/snes/ppu/sppu/sprite/sprite.cpp b/src/snes/ppu/sppu/sprite/sprite.cpp
index ba3c8ced..341ae433 100644
--- a/src/snes/ppu/sppu/sprite/sprite.cpp
+++ b/src/snes/ppu/sppu/sprite/sprite.cpp
@@ -26,136 +26,135 @@ void sPPU::Sprite::scanline() {
   auto oam_item = t.item[t.active];
   auto oam_tile = t.tile[t.active];
 
-  memset(oam_item, 0xff, 32);
-  for(unsigned i = 0; i < 34; i++) oam_tile[i].tile = 0xffff;
+  memset(oam_item, 0xff, 32);  //default to invalid
+  for(unsigned i = 0; i < 34; i++) oam_tile[i].x = 0xffff;  //default to invalid
 
   for(unsigned i = 0; i < 128; i++) {
-    t.active_sprite = (i + regs.first_sprite) & 127;
-    if(on_scanline() == false) continue;
+    unsigned sprite = (regs.first_sprite + i) & 127;
+    if(on_scanline(list[sprite]) == false) continue;
     if(t.item_count++ >= 32) break;
-    oam_item[t.item_count - 1] = (i + regs.first_sprite) & 127;
+    oam_item[t.item_count - 1] = sprite;
   }
 
   if(t.item_count > 0 && oam_item[t.item_count - 1] != 0xff) {
     ppu.regs.ioamaddr = 0x0200 + (oam_item[t.item_count - 1] >> 2);
   }
-
-  for(signed i = 31; i >= 0; i--) {
-    if(oam_item[i] == 0xff) continue;
-    t.active_sprite = oam_item[i];
-    load_tiles();
-  }
-
-  regs.time_over  |= (t.tile_count > 34);
-  regs.range_over |= (t.item_count > 32);
 }
 
-bool sPPU::Sprite::on_scanline() {
-  SpriteItem &sprite = list[t.active_sprite];
+bool sPPU::Sprite::on_scanline(SpriteItem &sprite) {
   if(sprite.x > 256 && (sprite.x + sprite.width() - 1) < 512) return false;
-
   signed height = (regs.interlace == false ? sprite.height() : (sprite.height() >> 1));
   if(t.y >= sprite.y && t.y < (sprite.y + height)) return true;
   if((sprite.y + height) >= 256 && t.y < ((sprite.y + height) & 255)) return true;
   return false;
 }
 
-void sPPU::Sprite::load_tiles() {
-  SpriteItem &sprite = list[t.active_sprite];
-  unsigned tile_width = sprite.width() >> 3;
-  signed x = sprite.x;
-  signed y = (t.y - sprite.y) & 0xff;
-  if(regs.interlace) y <<= 1;
-
-  if(sprite.vflip) {
-    if(sprite.width() == sprite.height()) {
-      y = (sprite.height() - 1) - y;
-    } else {
-      y = (y < sprite.width()) ? ((sprite.width() - 1) - y) : (sprite.width() + ((sprite.width() - 1) - (y - sprite.width())));
-    }
-  }
-
-  if(regs.interlace) {
-    y = (sprite.vflip == false ? y + self.field() : y - self.field());
-  }
-
-  x &= 511;
-  y &= 255;
-
-  uint16 tiledata_addr = regs.tiledata_addr;
-  uint16 chrx = (sprite.character >> 0) & 15;
-  uint16 chry = (sprite.character >> 4) & 15;
-  if(sprite.nameselect) {
-    tiledata_addr += (256 * 32) + (regs.nameselect << 13);
-  }
-  chry  += (y >> 3);
-  chry  &= 15;
-  chry <<= 4;
-
-  auto oam_tile = t.tile[t.active];
-
-  for(unsigned tx = 0; tx < tile_width; tx++) {
-    unsigned sx = (x + (tx << 3)) & 511;
-    if(x != 256 && sx >= 256 && (sx + 7) < 512) continue;
-
-    if(t.tile_count++ >= 34) break;
-    unsigned n = t.tile_count - 1;
-    oam_tile[n].x = sx;
-    oam_tile[n].y = y;
-    oam_tile[n].priority = sprite.priority;
-    oam_tile[n].palette = 128 + (sprite.palette << 4);
-    oam_tile[n].hflip = sprite.hflip;
-
-    unsigned mx = (sprite.hflip == false) ? tx : ((tile_width - 1) - tx);
-    unsigned pos = tiledata_addr + ((chry + ((chrx + mx) & 15)) << 5);
-    oam_tile[n].tile = (pos >> 5) & 0x07ff;
-  }
-}
-
 void sPPU::Sprite::run() {
   output.main.priority = 0;
   output.sub.priority = 0;
 
+  auto oam_tile = t.tile[!t.active];
   unsigned priority_table[] = { regs.priority0, regs.priority1, regs.priority2, regs.priority3 };
   unsigned x = t.x++;
 
   for(unsigned n = 0; n < 34; n++) {
-    TileItem &item = t.tile[!t.active][n];
-    if(item.tile == 0xffff) break;
+    auto tile = oam_tile[n];
+    if(tile.x == 0xffff) break;
 
-    int px = x - sclip<9>(item.x);
+    int px = x - sclip<9>(tile.x);
     if(px & ~7) continue;
 
-    uint16 addr = (item.tile << 5) + ((item.y & 7) * 2);
-    unsigned mask = 0x80 >> (item.hflip == false ? px : 7 - px);
-
-    uint8 d0 = memory::vram[addr +  0];
-    uint8 d1 = memory::vram[addr +  1];
-    uint8 d2 = memory::vram[addr + 16];
-    uint8 d3 = memory::vram[addr + 17];
-
+    unsigned mask = 0x80 >> (tile.hflip == false ? px : 7 - px);
     unsigned color;
-    color  = ((bool)(d0 & mask)) << 0;
-    color |= ((bool)(d1 & mask)) << 1;
-    color |= ((bool)(d2 & mask)) << 2;
-    color |= ((bool)(d3 & mask)) << 3;
+    color  = ((bool)(tile.d0 & mask)) << 0;
+    color |= ((bool)(tile.d1 & mask)) << 1;
+    color |= ((bool)(tile.d2 & mask)) << 2;
+    color |= ((bool)(tile.d3 & mask)) << 3;
 
     if(color) {
-      color += item.palette;
-
       if(regs.main_enabled) {
-        output.main.palette = color;
-        output.main.priority = priority_table[item.priority];
+        output.main.palette = tile.palette + color;
+        output.main.priority = priority_table[tile.priority];
       }
 
       if(regs.sub_enabled) {
-        output.sub.palette = color;
-        output.sub.priority = priority_table[item.priority];
+        output.sub.palette = tile.palette + color;
+        output.sub.priority = priority_table[tile.priority];
       }
     }
   }
 }
 
+void sPPU::Sprite::tilefetch() {
+  auto oam_item = t.item[t.active];
+  auto oam_tile = t.tile[t.active];
+
+  for(signed i = 31; i >= 0; i--) {
+    if(oam_item[i] == 0xff) continue;
+    auto sprite = list[oam_item[i]];
+
+    unsigned tile_width = sprite.width() >> 3;
+    signed x = sprite.x;
+    signed y = (t.y - sprite.y) & 0xff;
+    if(regs.interlace) y <<= 1;
+
+    if(sprite.vflip) {
+      if(sprite.width() == sprite.height()) {
+        y = (sprite.height() - 1) - y;
+      } else if(y < sprite.width()) {
+        y = (sprite.width() - 1) - y;
+      } else {
+        y = sprite.width() + ((sprite.width() - 1) - (y - sprite.width()));
+      }
+    }
+
+    if(regs.interlace) {
+      y = (sprite.vflip == false ? y + self.field() : y - self.field());
+    }
+
+    x &= 511;
+    y &= 255;
+
+    uint16 tiledata_addr = regs.tiledata_addr;
+    uint16 chrx = (sprite.character >> 0) & 15;
+    uint16 chry = (sprite.character >> 4) & 15;
+    if(sprite.nameselect) {
+      tiledata_addr += (256 * 32) + (regs.nameselect << 13);
+    }
+    chry  += (y >> 3);
+    chry  &= 15;
+    chry <<= 4;
+
+    for(unsigned tx = 0; tx < tile_width; tx++) {
+      unsigned sx = (x + (tx << 3)) & 511;
+      if(x != 256 && sx >= 256 && (sx + 7) < 512) continue;
+      if(t.tile_count++ >= 34) break;
+
+      unsigned n = t.tile_count - 1;
+      oam_tile[n].x = sx;
+      oam_tile[n].priority = sprite.priority;
+      oam_tile[n].palette = 128 + (sprite.palette << 4);
+      oam_tile[n].hflip = sprite.hflip;
+
+      unsigned mx = (sprite.hflip == false) ? tx : ((tile_width - 1) - tx);
+      unsigned pos = tiledata_addr + ((chry + ((chrx + mx) & 15)) << 5);
+      uint16 addr = (pos & 0xffe0) + ((y & 7) * 2);
+
+      oam_tile[n].d0 = memory::vram[addr +  0];
+      oam_tile[n].d1 = memory::vram[addr +  1];
+      self.add_clocks(2);
+
+      oam_tile[n].d2 = memory::vram[addr + 16];
+      oam_tile[n].d3 = memory::vram[addr + 17];
+      self.add_clocks(2);
+    }
+  }
+
+  if(t.tile_count < 34) self.add_clocks((34 - t.tile_count) * 4);
+  regs.time_over  |= (t.tile_count > 34);
+  regs.range_over |= (t.item_count > 32);
+}
+
 void sPPU::Sprite::reset() {
   for(unsigned i = 0; i < 128; i++) {
     list[i].x = 0;
@@ -172,7 +171,6 @@ void sPPU::Sprite::reset() {
   t.x = 0;
   t.y = 0;
 
-  t.active_sprite = 0;
   t.item_count = 0;
   t.tile_count = 0;
 
@@ -181,11 +179,13 @@ void sPPU::Sprite::reset() {
     memset(t.item[n], 0, 32);
     for(unsigned i = 0; i < 34; i++) {
       t.tile[n][i].x = 0;
-      t.tile[n][i].y = 0;
       t.tile[n][i].priority = 0;
       t.tile[n][i].palette = 0;
-      t.tile[n][i].tile = 0;
       t.tile[n][i].hflip = 0;
+      t.tile[n][i].d0 = 0;
+      t.tile[n][i].d1 = 0;
+      t.tile[n][i].d2 = 0;
+      t.tile[n][i].d3 = 0;
     }
   }
 
diff --git a/src/snes/ppu/sppu/sprite/sprite.hpp b/src/snes/ppu/sppu/sprite/sprite.hpp
index daa77b4e..2bf95711 100644
--- a/src/snes/ppu/sppu/sprite/sprite.hpp
+++ b/src/snes/ppu/sppu/sprite/sprite.hpp
@@ -18,18 +18,16 @@ public:
 
   struct TileItem {
     uint16 x;
-    uint16 y;
     uint16 priority;
     uint16 palette;
-    uint16 tile;
     bool hflip;
+    uint8 d0, d1, d2, d3;
   };
 
   struct State {
     unsigned x;
     unsigned y;
 
-    unsigned active_sprite;
     unsigned item_count;
     unsigned tile_count;
 
@@ -72,12 +70,12 @@ public:
   void frame();
   void scanline();
   void run();
+  void tilefetch();
   void reset();
 
   void serialize(serializer&);
   Sprite(sPPU &self);
 
 private:
-  bool on_scanline();
-  void load_tiles();
+  bool on_scanline(SpriteItem&);
 };
diff --git a/src/snes/snes.hpp b/src/snes/snes.hpp
index 96baa8a4..0b4da2cb 100644
--- a/src/snes/snes.hpp
+++ b/src/snes/snes.hpp
@@ -1,4 +1,4 @@
-static const char bsnesVersion[] = "067.03";
+static const char bsnesVersion[] = "067.04";
 static const char bsnesTitle[] = "bsnes";
 static const unsigned bsnesSerializerVersion = 12;