diff --git a/source/font.cpp b/source/font.cpp
index b4b5825..63fb1ed 100644
--- a/source/font.cpp
+++ b/source/font.cpp
@@ -216,6 +216,9 @@ ITCM_CODE void Font::print(std::u16string_view text, int x, int y, bool top, int
 	}
 	const int xStart = x;
 
+	if(maxWidth != 0)
+		scaleX = std::min(scaleX, (float)maxWidth / (calcWidth(text) * scaleX));
+
 	bool rtl = false;
 	for(const auto c : text) {
 		if(c >= 0x0590 && c <= 0x05FF) {
@@ -273,7 +276,7 @@ ITCM_CODE void Font::print(std::u16string_view text, int x, int y, bool top, int
 
 		if(*it == '\n') {
 			x = xStart;
-			y += tileHeight;
+			y += tileHeight * scaleY;
 			continue;
 		}
 
@@ -322,44 +325,44 @@ ITCM_CODE void Font::print(std::u16string_view text, int x, int y, bool top, int
 			// Don't draw off screen chars
 			if(x >= 0 && x + fontWidths[(index * 3) + 2] < 256 && y >= 0 && y + tileHeight < 192) {
 #ifdef TEXT_BUFFERED
-				u8 *dst = textBuf[top] + x + fontWidths[(index * 3)];
+				u8 *dst = textBuf[top] + y * 256 + x + fontWidths[(index * 3)];
 #else
-				u8 *dst = (u8 *)bgGetGfxPtr(top ? layer : layer + 4) + x + fontWidths[(index * 3)];
+				u8 *dst = (u8 *)bgGetGfxPtr(top ? layer : layer + 4) + y * 256 + x + fontWidths[(index * 3)];
 #endif
 				// Use faster integer math if scale is 1
 				if(scaleX == 1.0f && scaleY == 1.0f) {
 					for(int i = 0; i < tileHeight; i++) {
 						for(int j = 0; j < tileWidth; j++) {
 							u8 px = fontTiles[(index * tileSize) + (i * tileWidth + j) / 4] >>
-									((3 - ((i * tileWidth + j) % 4)) * 2) &
-								3;
-							if(px)
+									((3 - ((i * tileWidth + j) % 4)) * 2) & 3;
+							if(px) {
 #ifdef TEXT_BUFFERED
-								dst[(y + i) * 256 + j] = px + (color * 4);
+								dst[i * 256 + j] = px + (color * 4);
 #else
-								toncset(dst + (y + i) * 256 + j, px + (color * 4), 1);
+								toncset(dst + i * 256 + j, px + (color * 4), 1);
 #endif
+							}
 						}
 					}
 				} else {
-					for(float i = 0.0f; i < tileHeight; i += 1 / scaleY) {
-						for(float j = 0.0f; j < tileWidth; j += 1 / scaleX) {
-							u8 px = fontTiles[(index * tileSize) + int(i * tileWidth + j) / 4] >>
-									((3 - (int(i * tileWidth + j) % 4)) * 2) &
-								3;
-							if(px)
+					for(int i = 0; i < tileHeight * scaleY; i++) {
+						for(int j = 0; j < tileWidth * scaleX; j++) {
+							u8 loc = int(i / scaleY) * tileWidth + int(j / scaleX);
+							u8 px = fontTiles[index * tileSize + loc / 4] >> ((3 - (loc % 4)) * 2) & 3;
+							if(px) {
 #ifdef TEXT_BUFFERED
-								dst[int((y + i) * 256 + j)] = px + (color * 4);
+								dst[i * 256 + j] = px + (color * 4);
 #else
-								toncset(dst + int((y + i) * 256 + j), px + (color * 4), 1);
+								toncset(dst + i * 256 + j, px + (color * 4), 1);
 #endif
+							}
 						}
 					}
 				}
 			}
 		}
 
-		x += fontWidths[(index * 3) + 2];
+		x += fontWidths[(index * 3) + 2] * scaleX;
 	}
 }
 
diff --git a/source/image.cpp b/source/image.cpp
index 73a5c5a..c2713fb 100644
--- a/source/image.cpp
+++ b/source/image.cpp
@@ -93,10 +93,10 @@ void Image::draw(int x, int y, bool top, int layer, int channel, bool copyPal) {
 
 	// If full width and X is 0, copy it all in one go
 	if(_width == width && x == 0) {
-		dmaCopyHalfWords(channel, _bitmap.data(), dst, _width * _height);
+		tonccpy(dst, _bitmap.data(), _width * _height);
 	} else {
 		for(int i = 0; i < _height; i++) {
-			dmaCopyHalfWords(channel, _bitmap.data() + (i * _width), dst + (y + i) * 256 + x, _width);
+			tonccpy(dst + (y + i) * 256 + x, _bitmap.data() + i * _width, _width);
 		}
 	}
 }
@@ -106,7 +106,7 @@ void Image::drawSpecial(int x, int y, bool top, int layer, float scaleX, float s
 	if(copyPal)
 		tonccpy((top ? BG_PALETTE : BG_PALETTE_SUB) + _palOfs + paletteOffset, _palette.data(), _palette.size() * 2);
 
-	u8 *dst = (u8 *)bgGetGfxPtr(top ? layer : layer + 4);
+	u8 *dst = (u8 *)bgGetGfxPtr(top ? layer : layer + 4) + y * 256 + x;
 
 	// If the scale is 1 use faster integer math
 	if(scaleX == 1.0f && scaleY == 1.0f) {
@@ -114,15 +114,15 @@ void Image::drawSpecial(int x, int y, bool top, int layer, float scaleX, float s
 			for(float j = 0; j < _width; j++) {
 				u8 px = _bitmap[i * _width + j];
 				if(_palette[px - _palOfs] & 0x8000)
-					toncset(dst + int((y + i) * 256 + x + j), px + paletteOffset, 1);
+					toncset(dst + i * 256 + x, px + paletteOffset, 1);
 			}
 		}
 	} else {
-		for(float i = 0.0f; i < _height; i += 1 / scaleY) {
-			for(float j = 0.0f; j < _width; j += 1 / scaleX) {
-				u8 px = _bitmap[int((i * _width) + j)];
+		for(int i = 0; i < _height * scaleY; i++) {
+			for(int j = 0; j < _width * scaleX; j++) {
+				u8 px = _bitmap[int(i / scaleY) * _width + int(j / scaleX)];
 				if(_palette[px - _palOfs] & 0x8000)
-					toncset(dst + int((y + i) * 256 + x + j), px + paletteOffset, 1);
+					toncset(dst + i * 256 + j, px + paletteOffset, 1);
 			}
 		}
 	}
@@ -134,8 +134,7 @@ void Image::drawSegment(int x, int y, int imageX, int imageY, int w, int h, bool
 		tonccpy((top ? BG_PALETTE : BG_PALETTE_SUB) + _palOfs, _palette.data(), _palette.size() * 2);
 
 	for(int i = 0; i < h; i++) {
-		dmaCopyHalfWords(channel, _bitmap.data() + ((imageY + i) * _width) + imageX,
-						 (u8 *)bgGetGfxPtr(top ? layer : layer + 4) + ((y + i) * 256) + x, w);
+		tonccpy((u8 *)bgGetGfxPtr(top ? layer : layer + 4) + (y + i) * 256 + x, _bitmap.data() + (imageY + i) * _width + imageX, w);
 	}
 }
 
@@ -144,7 +143,7 @@ void Image::drawSegmentSpecial(int x, int y, int imageX, int imageY, int w, int
 	if(copyPal)
 		tonccpy((top ? BG_PALETTE : BG_PALETTE_SUB) + _palOfs + paletteOffset, _palette.data(), _palette.size() * 2);
 
-	u8 *dst = (u8 *)bgGetGfxPtr(top ? layer : layer + 4);
+	u8 *dst = (u8 *)bgGetGfxPtr(top ? layer : layer + 4) + y * 256 + x;
 
 	// If the scale is 1 use faster integer math
 	if(scaleX == 1.0f && scaleY == 1.0f) {
@@ -152,15 +151,15 @@ void Image::drawSegmentSpecial(int x, int y, int imageX, int imageY, int w, int
 			for(int j = 0; j < w; j++) {
 				u8 px = _bitmap[i * _width + j];
 				if(_palette[px - _palOfs] & 0x8000)
-					toncset(dst + ((y + i) * 256 + x + j), px + paletteOffset, 1);
+					toncset(dst + i * 256 + j, px + paletteOffset, 1);
 			}
 		}
 	} else {
-		for(float i = 0; i < h; i += 1 / scaleX) {
-			for(float j = 0; j < _width; j += 1 / scaleY) {
-				u8 px = _bitmap[i * _width + j];
+		for(int i = 0; i < h * scaleY; i++) {
+			for(int j = 0; j < w * scaleX; j++) {
+				u8 px = _bitmap[(imageY + int(i / scaleY)) * _width + imageX + int(j / scaleX)];
 				if(_palette[px - _palOfs] & 0x8000)
-					toncset(dst + int((y + i) * 256 + x + j), px + paletteOffset, 1);
+					toncset(dst + i * 256 + j, px + paletteOffset, 1);
 			}
 		}
 	}
diff --git a/source/sprite.cpp b/source/sprite.cpp
index 9b592fa..f08146e 100644
--- a/source/sprite.cpp
+++ b/source/sprite.cpp
@@ -159,11 +159,11 @@ void Sprite::drawImage(int x, int y, const Image &image, float scaleX, float sca
 			}
 		}
 	} else {
-		for(float i = 0.0f; i < image.height(); i += 1 / scaleY) {
-			for(float j = 0.0f; j < image.width(); j += 1 / scaleX) {
-				u16 px = image.palette()[image.bitmap()[int(i * image.width() + j)] - image.palOfs()];
+		for(int i = 0; i < image.height() * scaleY; i++) {
+			for(int j = 0; j < image.width() * scaleX; j++) {
+				u16 px = image.palette()[image.bitmap()[int(i / scaleY) * image.width() + int(j / scaleX)] - image.palOfs()];
 				if(px & 0x8000)
-					toncset16(_gfx + int((y + i) * _height + j) + x, px, 1);
+					toncset16(_gfx + (y + i) * _height + j + x, px, 1);
 			}
 		}
 	}
@@ -181,11 +181,11 @@ void Sprite::drawImageSegment(int x, int y, int imageX, int imageY, int w, int h
 			}
 		}
 	} else {
-		for(float i = 0; i < h; i += 1 / scaleX) {
-			for(float j = 0; j < image.width(); j += 1 / scaleY) {
-				u16 px = image.palette()[image.bitmap()[(imageY + i) * image.width() + imageX + j] - image.palOfs()];
+		for(int i = 0; i < h * scaleY; i++) {
+			for(int j = 0; j < w * scaleX; j++) {
+				u16 px = image.palette()[image.bitmap()[(imageY + int(i / scaleY)) * image.width() + imageX + int(j / scaleX)] - image.palOfs()];
 				if(px & 0x8000)
-					toncset16(_gfx + int((y + i) * _height + x + j), px, 1);
+					toncset16(_gfx + (y + i) * _height + x + j, px, 1);
 			}
 		}
 	}