diff --git a/src/pkg/exp/draw/color.go b/src/pkg/exp/draw/color.go
index f0e7597092b..3fe7b4abc7c 100644
--- a/src/pkg/exp/draw/color.go
+++ b/src/pkg/exp/draw/color.go
@@ -51,13 +51,9 @@ func (c Color) RGBA() (r, g, b, a uint32) {
 	x := uint32(c)
 	r, g, b, a = x>>24, (x>>16)&0xFF, (x>>8)&0xFF, x&0xFF
 	r |= r << 8
-	r |= r << 16
 	g |= g << 8
-	g |= g << 16
 	b |= b << 8
-	b |= b << 16
 	a |= a << 8
-	a |= a << 16
 	return
 }
 
@@ -103,7 +99,7 @@ func toColor(color image.Color) image.Color {
 		return c
 	}
 	r, g, b, a := color.RGBA()
-	return Color(r>>24<<24 | g>>24<<16 | b>>24<<8 | a>>24)
+	return Color(r>>8<<24 | g>>8<<16 | b>>8<<8 | a>>8)
 }
 
 func (c Color) ColorModel() image.ColorModel { return image.ColorModelFunc(toColor) }
diff --git a/src/pkg/exp/draw/draw.go b/src/pkg/exp/draw/draw.go
index 41eaef4d4f3..d9d3d13eb09 100644
--- a/src/pkg/exp/draw/draw.go
+++ b/src/pkg/exp/draw/draw.go
@@ -105,13 +105,10 @@ func DrawMask(dst Image, r Rectangle, src image.Image, sp Point, mask image.Imag
 		sx := sp.X + x0 - r.Min.X
 		mx := mp.X + x0 - r.Min.X
 		for x := x0; x != x1; x, sx, mx = x+dx, sx+dx, mx+dx {
-			// A nil mask is equivalent to a fully opaque, infinitely large mask.
-			// We work in 16-bit color, so that multiplying two values does not overflow a uint32.
 			const M = 1<<16 - 1
 			ma := uint32(M)
 			if mask != nil {
 				_, _, _, ma = mask.At(mx, my).RGBA()
-				ma >>= 16
 			}
 			switch {
 			case ma == 0:
@@ -124,19 +121,11 @@ func DrawMask(dst Image, r Rectangle, src image.Image, sp Point, mask image.Imag
 				dst.Set(x, y, src.At(sx, sy))
 			default:
 				sr, sg, sb, sa := src.At(sx, sy).RGBA()
-				sr >>= 16
-				sg >>= 16
-				sb >>= 16
-				sa >>= 16
 				if out == nil {
 					out = new(image.RGBA64Color)
 				}
 				if op == Over {
 					dr, dg, db, da := dst.At(x, y).RGBA()
-					dr >>= 16
-					dg >>= 16
-					db >>= 16
-					da >>= 16
 					a := M - (sa * ma / M)
 					out.R = uint16((dr*a + sr*ma) / M)
 					out.G = uint16((dg*a + sg*ma) / M)
@@ -158,10 +147,6 @@ func drawGlyphOver(dst *image.RGBA, r Rectangle, src image.ColorImage, mask *ima
 	x0, x1 := r.Min.X, r.Max.X
 	y0, y1 := r.Min.Y, r.Max.Y
 	cr, cg, cb, ca := src.RGBA()
-	cr >>= 16
-	cg >>= 16
-	cb >>= 16
-	ca >>= 16
 	for y, my := y0, mp.Y; y != y1; y, my = y+1, my+1 {
 		p := dst.Pixel[y]
 		for x, mx := x0, mp.X; x != x1; x, mx = x+1, mx+1 {
@@ -192,7 +177,7 @@ func drawFill(dst *image.RGBA, r Rectangle, src image.ColorImage) {
 		return
 	}
 	cr, cg, cb, ca := src.RGBA()
-	color := image.RGBAColor{uint8(cr >> 24), uint8(cg >> 24), uint8(cb >> 24), uint8(ca >> 24)}
+	color := image.RGBAColor{uint8(cr >> 8), uint8(cg >> 8), uint8(cb >> 8), uint8(ca >> 8)}
 	// The built-in copy function is faster than a straightforward for loop to fill the destination with
 	// the color, but copy requires a slice source. We therefore use a for loop to fill the first row, and
 	// then use the first row as the slice source for the remaining rows.
@@ -238,13 +223,8 @@ func drawRGBA(dst *image.RGBA, r Rectangle, src image.Image, sp Point, mask imag
 			ma := uint32(M)
 			if mask != nil {
 				_, _, _, ma = mask.At(mx, my).RGBA()
-				ma >>= 16
 			}
 			sr, sg, sb, sa := src.At(sx, sy).RGBA()
-			sr >>= 16
-			sg >>= 16
-			sb >>= 16
-			sa >>= 16
 			var dr, dg, db, da uint32
 			if op == Over {
 				rgba := p[x]
diff --git a/src/pkg/exp/draw/draw_test.go b/src/pkg/exp/draw/draw_test.go
index 675c4eaec5f..5303f2b3d85 100644
--- a/src/pkg/exp/draw/draw_test.go
+++ b/src/pkg/exp/draw/draw_test.go
@@ -106,20 +106,11 @@ func makeGolden(dst image.Image, t drawTest) image.Image {
 			var dr, dg, db, da uint32
 			if t.op == Over {
 				dr, dg, db, da = dst.At(x, y).RGBA()
-				dr >>= 16
-				dg >>= 16
-				db >>= 16
-				da >>= 16
 			}
 			sr, sg, sb, sa := t.src.At(sx, sy).RGBA()
-			sr >>= 16
-			sg >>= 16
-			sb >>= 16
-			sa >>= 16
 			ma := uint32(M)
 			if t.mask != nil {
 				_, _, _, ma = t.mask.At(mx, my).RGBA()
-				ma >>= 16
 			}
 			a := M - (sa * ma / M)
 			golden.Set(x, y, image.RGBA64Color{
diff --git a/src/pkg/exp/nacl/av/image.go b/src/pkg/exp/nacl/av/image.go
index f69f94a657e..8de5311365c 100644
--- a/src/pkg/exp/nacl/av/image.go
+++ b/src/pkg/exp/nacl/av/image.go
@@ -61,16 +61,12 @@ func (p Color) RGBA() (r, g, b, a uint32) {
 	x := uint32(p)
 	a = x >> 24
 	a |= a << 8
-	a |= a << 16
 	r = (x >> 16) & 0xFF
 	r |= r << 8
-	r |= r << 16
 	g = (x >> 8) & 0xFF
 	g |= g << 8
-	g |= g << 16
 	b = x & 0xFF
 	b |= b << 8
-	b |= b << 16
 	return
 }
 
diff --git a/src/pkg/image/color.go b/src/pkg/image/color.go
index 31ba59280f4..c17ffc38946 100644
--- a/src/pkg/image/color.go
+++ b/src/pkg/image/color.go
@@ -4,14 +4,15 @@
 
 package image
 
-// TODO(nigeltao): Think about how floating-point color models work.
-
-// All Colors can convert themselves, with a possible loss of precision, to 128-bit alpha-premultiplied RGBA.
+// All Colors can convert themselves, with a possible loss of precision,
+// to 64-bit alpha-premultiplied RGBA. Each channel value ranges within
+// [0, 0xFFFF].
 type Color interface {
 	RGBA() (r, g, b, a uint32)
 }
 
-// An RGBAColor represents a traditional 32-bit alpha-premultiplied color, having 8 bits for each of red, green, blue and alpha.
+// An RGBAColor represents a traditional 32-bit alpha-premultiplied color,
+// having 8 bits for each of red, green, blue and alpha.
 type RGBAColor struct {
 	R, G, B, A uint8
 }
@@ -19,34 +20,23 @@ type RGBAColor struct {
 func (c RGBAColor) RGBA() (r, g, b, a uint32) {
 	r = uint32(c.R)
 	r |= r << 8
-	r |= r << 16
 	g = uint32(c.G)
 	g |= g << 8
-	g |= g << 16
 	b = uint32(c.B)
 	b |= b << 8
-	b |= b << 16
 	a = uint32(c.A)
 	a |= a << 8
-	a |= a << 16
 	return
 }
 
-// An RGBA64Color represents a 64-bit alpha-premultiplied color, having 16 bits for each of red, green, blue and alpha.
+// An RGBA64Color represents a 64-bit alpha-premultiplied color,
+// having 16 bits for each of red, green, blue and alpha.
 type RGBA64Color struct {
 	R, G, B, A uint16
 }
 
 func (c RGBA64Color) RGBA() (r, g, b, a uint32) {
-	r = uint32(c.R)
-	r |= r << 16
-	g = uint32(c.G)
-	g |= g << 16
-	b = uint32(c.B)
-	b |= b << 16
-	a = uint32(c.A)
-	a |= a << 16
-	return
+	return uint32(c.R), uint32(c.G), uint32(c.B), uint32(c.A)
 }
 
 // An NRGBAColor represents a non-alpha-premultiplied 32-bit color.
@@ -59,24 +49,21 @@ func (c NRGBAColor) RGBA() (r, g, b, a uint32) {
 	r |= r << 8
 	r *= uint32(c.A)
 	r /= 0xff
-	r |= r << 16
 	g = uint32(c.G)
 	g |= g << 8
 	g *= uint32(c.A)
 	g /= 0xff
-	g |= g << 16
 	b = uint32(c.B)
 	b |= b << 8
 	b *= uint32(c.A)
 	b /= 0xff
-	b |= b << 16
 	a = uint32(c.A)
 	a |= a << 8
-	a |= a << 16
 	return
 }
 
-// An NRGBA64Color represents a non-alpha-premultiplied 64-bit color, having 16 bits for each of red, green, blue and alpha.
+// An NRGBA64Color represents a non-alpha-premultiplied 64-bit color,
+// having 16 bits for each of red, green, blue and alpha.
 type NRGBA64Color struct {
 	R, G, B, A uint16
 }
@@ -85,18 +72,13 @@ func (c NRGBA64Color) RGBA() (r, g, b, a uint32) {
 	r = uint32(c.R)
 	r *= uint32(c.A)
 	r /= 0xffff
-	r |= r << 16
 	g = uint32(c.G)
 	g *= uint32(c.A)
 	g /= 0xffff
-	g |= g << 16
 	b = uint32(c.B)
 	b *= uint32(c.A)
 	b /= 0xffff
-	b |= b << 16
 	a = uint32(c.A)
-	a |= a << 8
-	a |= a << 16
 	return
 }
 
@@ -108,12 +90,11 @@ type AlphaColor struct {
 func (c AlphaColor) RGBA() (r, g, b, a uint32) {
 	a = uint32(c.A)
 	a |= a << 8
-	a |= a << 16
 	return a, a, a, a
 }
 
-// A ColorModel can convert foreign Colors, with a possible loss of precision, to a Color
-// from its own color model.
+// A ColorModel can convert foreign Colors, with a possible loss of precision,
+// to a Color from its own color model.
 type ColorModel interface {
 	Convert(c Color) Color
 }
@@ -129,36 +110,32 @@ func (f ColorModelFunc) Convert(c Color) Color {
 }
 
 func toRGBAColor(c Color) Color {
-	if _, ok := c.(RGBAColor); ok { // no-op conversion
+	if _, ok := c.(RGBAColor); ok {
 		return c
 	}
 	r, g, b, a := c.RGBA()
-	return RGBAColor{uint8(r >> 24), uint8(g >> 24), uint8(b >> 24), uint8(a >> 24)}
+	return RGBAColor{uint8(r >> 8), uint8(g >> 8), uint8(b >> 8), uint8(a >> 8)}
 }
 
 func toRGBA64Color(c Color) Color {
-	if _, ok := c.(RGBA64Color); ok { // no-op conversion
+	if _, ok := c.(RGBA64Color); ok {
 		return c
 	}
 	r, g, b, a := c.RGBA()
-	return RGBA64Color{uint16(r >> 16), uint16(g >> 16), uint16(b >> 16), uint16(a >> 16)}
+	return RGBA64Color{uint16(r), uint16(g), uint16(b), uint16(a)}
 }
 
 func toNRGBAColor(c Color) Color {
-	if _, ok := c.(NRGBAColor); ok { // no-op conversion
+	if _, ok := c.(NRGBAColor); ok {
 		return c
 	}
 	r, g, b, a := c.RGBA()
-	a >>= 16
 	if a == 0xffff {
-		return NRGBAColor{uint8(r >> 24), uint8(g >> 24), uint8(b >> 24), 0xff}
+		return NRGBAColor{uint8(r >> 8), uint8(g >> 8), uint8(b >> 8), 0xff}
 	}
 	if a == 0 {
 		return NRGBAColor{0, 0, 0, 0}
 	}
-	r >>= 16
-	g >>= 16
-	b >>= 16
 	// Since Color.RGBA returns a alpha-premultiplied color, we should have r <= a && g <= a && b <= a.
 	r = (r * 0xffff) / a
 	g = (g * 0xffff) / a
@@ -167,14 +144,10 @@ func toNRGBAColor(c Color) Color {
 }
 
 func toNRGBA64Color(c Color) Color {
-	if _, ok := c.(NRGBA64Color); ok { // no-op conversion
+	if _, ok := c.(NRGBA64Color); ok {
 		return c
 	}
 	r, g, b, a := c.RGBA()
-	a >>= 16
-	r >>= 16
-	g >>= 16
-	b >>= 16
 	if a == 0xffff {
 		return NRGBA64Color{uint16(r), uint16(g), uint16(b), 0xffff}
 	}
@@ -189,11 +162,11 @@ func toNRGBA64Color(c Color) Color {
 }
 
 func toAlphaColor(c Color) Color {
-	if _, ok := c.(AlphaColor); ok { // no-op conversion
+	if _, ok := c.(AlphaColor); ok {
 		return c
 	}
 	_, _, _, a := c.RGBA()
-	return AlphaColor{uint8(a >> 24)}
+	return AlphaColor{uint8(a >> 8)}
 }
 
 // The ColorModel associated with RGBAColor.
diff --git a/src/pkg/image/image.go b/src/pkg/image/image.go
index f4ac823e1be..3ac7d4eb2a8 100644
--- a/src/pkg/image/image.go
+++ b/src/pkg/image/image.go
@@ -185,21 +185,18 @@ func (p PalettedColorModel) Convert(c Color) Color {
 	if len(p) == 0 {
 		return nil
 	}
-	// TODO(nigeltao): Revisit the "pick the palette color which minimizes sum-squared-difference"
-	// algorithm when the premultiplied vs unpremultiplied issue is resolved.
-	// Currently, we only compare the R, G and B values, and ignore A.
 	cr, cg, cb, _ := c.RGBA()
-	// Shift by 17 bits to avoid potential uint32 overflow in sum-squared-difference.
-	cr >>= 17
-	cg >>= 17
-	cb >>= 17
+	// Shift by 1 bit to avoid potential uint32 overflow in sum-squared-difference.
+	cr >>= 1
+	cg >>= 1
+	cb >>= 1
 	result := Color(nil)
 	bestSSD := uint32(1<<32 - 1)
 	for _, v := range p {
 		vr, vg, vb, _ := v.RGBA()
-		vr >>= 17
-		vg >>= 17
-		vb >>= 17
+		vr >>= 1
+		vg >>= 1
+		vb >>= 1
 		dr, dg, db := diff(cr, vr), diff(cg, vg), diff(cb, vb)
 		ssd := (dr * dr) + (dg * dg) + (db * db)
 		if ssd < bestSSD {
diff --git a/src/pkg/image/png/reader_test.go b/src/pkg/image/png/reader_test.go
index 68aab78322c..1dc45992e93 100644
--- a/src/pkg/image/png/reader_test.go
+++ b/src/pkg/image/png/reader_test.go
@@ -76,9 +76,9 @@ func sng(w io.WriteCloser, filename string, png image.Image) {
 		io.WriteString(w, "PLTE {\n")
 		for i := 0; i < len(cpm); i++ {
 			r, g, b, _ := cpm[i].RGBA()
-			r >>= 24
-			g >>= 24
-			b >>= 24
+			r >>= 8
+			g >>= 8
+			b >>= 8
 			fmt.Fprintf(w, "    (%3d,%3d,%3d)     # rgb = (0x%02x,0x%02x,0x%02x)\n", r, g, b, r, g, b)
 		}
 		io.WriteString(w, "}\n")
diff --git a/src/pkg/image/png/writer.go b/src/pkg/image/png/writer.go
index 06b6dcdc338..e186ca819b9 100644
--- a/src/pkg/image/png/writer.go
+++ b/src/pkg/image/png/writer.go
@@ -37,7 +37,7 @@ func opaque(m image.Image) bool {
 	for y := 0; y < m.Height(); y++ {
 		for x := 0; x < m.Width(); x++ {
 			_, _, _, a := m.At(x, y).RGBA()
-			if a != 0xffffffff {
+			if a != 0xffff {
 				return false
 			}
 		}
@@ -101,13 +101,13 @@ func (e *encoder) writePLTE(p image.PalettedColorModel) {
 	}
 	for i := 0; i < len(p); i++ {
 		r, g, b, a := p[i].RGBA()
-		if a != 0xffffffff {
+		if a != 0xffff {
 			e.err = UnsupportedError("non-opaque palette color")
 			return
 		}
-		e.tmp[3*i+0] = uint8(r >> 24)
-		e.tmp[3*i+1] = uint8(g >> 24)
-		e.tmp[3*i+2] = uint8(b >> 24)
+		e.tmp[3*i+0] = uint8(r >> 8)
+		e.tmp[3*i+1] = uint8(g >> 8)
+		e.tmp[3*i+2] = uint8(b >> 8)
 	}
 	e.writeChunk(e.tmp[0:3*len(p)], "PLTE")
 }
@@ -261,9 +261,9 @@ func writeImage(w io.Writer, m image.Image, ct uint8) os.Error {
 			for x := 0; x < m.Width(); x++ {
 				// We have previously verified that the alpha value is fully opaque.
 				r, g, b, _ := m.At(x, y).RGBA()
-				cr[0][3*x+1] = uint8(r >> 24)
-				cr[0][3*x+2] = uint8(g >> 24)
-				cr[0][3*x+3] = uint8(b >> 24)
+				cr[0][3*x+1] = uint8(r >> 8)
+				cr[0][3*x+2] = uint8(g >> 8)
+				cr[0][3*x+3] = uint8(b >> 8)
 			}
 		case ctPaletted:
 			for x := 0; x < m.Width(); x++ {