diff --git a/libvips/colour/LCh2Lab.c b/libvips/colour/LCh2Lab.c
index 9f615cd3..861874a3 100644
--- a/libvips/colour/LCh2Lab.c
+++ b/libvips/colour/LCh2Lab.c
@@ -57,6 +57,22 @@ typedef VipsColourTransformClass VipsLCh2LabClass;
 
 G_DEFINE_TYPE( VipsLCh2Lab, vips_LCh2Lab, VIPS_TYPE_COLOUR_TRANSFORM );
 
+/**
+ * vips_col_Ch2ab:
+ * @C: Chroma
+ * @h: Hue angle (degrees)
+ * @a: return CIE a* value
+ * @b: return CIE b* value
+ *
+ * Calculate ab from Ch, h in degrees.
+ */
+void
+vips_col_Ch2ab( float C, float h, float *a, float *b )
+{
+	*a = C * cos( VIPS_RAD( h ) );
+	*b = C * sin( VIPS_RAD( h ) );
+}
+
 /* Process a buffer of data.
  */
 static void
@@ -75,8 +91,7 @@ vips_LCh2Lab_line( VipsColour *colour, VipsPel *out, VipsPel **in, int width )
 
 		p += 3;
 
-		a = C * cos( VIPS_RAD( h ) );
-		b = C * sin( VIPS_RAD( h ) );
+		vips_col_Ch2ab( C, h, &a, &b );
 
 		q[0] = L;
 		q[1] = a;
@@ -86,35 +101,6 @@ vips_LCh2Lab_line( VipsColour *colour, VipsPel *out, VipsPel **in, int width )
 	}
 }
 
-/**
- * vips_col_Ch2ab:
- * @C: Chroma
- * @h: Hue angle (degrees)
- * @a: return CIE a* value
- * @b: return CIE b* value
- *
- * Calculate ab from Ch, h in degrees.
- */
-void
-vips_col_Ch2ab( float C, float h, float *a, float *b )
-{
-	float in[3];
-	float out[3];
-	float *x;
-
-	/* could be anything, we don't use this value, but we must supply one
-	 * or static analyzers will complain.
-	 */
-	in[0] = 50.0;
-
-	in[1] = C;
-	in[2] = h;
-	x = &in[0];
-	vips_LCh2Lab_line( NULL, (VipsPel *) out, (VipsPel **) &x, 1 );
-	*a = out[1];
-	*b = out[2];
-}
-
 static void
 vips_LCh2Lab_class_init( VipsLCh2LabClass *class )
 {
diff --git a/libvips/colour/Lab2XYZ.c b/libvips/colour/Lab2XYZ.c
index 5e229e41..2d65d2e6 100644
--- a/libvips/colour/Lab2XYZ.c
+++ b/libvips/colour/Lab2XYZ.c
@@ -80,6 +80,34 @@ typedef VipsColourTransformClass VipsLab2XYZClass;
 
 G_DEFINE_TYPE( VipsLab2XYZ, vips_Lab2XYZ, VIPS_TYPE_COLOUR_TRANSFORM );
 
+static void
+vips_col_Lab2XYZ_helper( VipsLab2XYZ *Lab2XYZ,
+	float L, float a, float b, float *X, float *Y, float *Z )
+{
+	double cby, tmp;
+
+	if( L < 8.0 ) {
+		*Y = (L * Lab2XYZ->Y0) / 903.3;
+		cby = 7.787 * (*Y / Lab2XYZ->Y0) + 16.0 / 116.0;
+	}
+	else {
+		cby = (L + 16.0) / 116.0;
+		*Y = Lab2XYZ->Y0 * cby * cby * cby;
+	}
+
+	tmp = a / 500.0 + cby;
+	if( tmp < 0.2069 )
+		*X = Lab2XYZ->X0 * (tmp - 0.13793) / 7.787;
+	else
+		*X = Lab2XYZ->X0 * tmp * tmp * tmp;
+
+	tmp = cby - b / 200.0;
+	if( tmp < 0.2069 )
+		*Z = Lab2XYZ->Z0 * (tmp - 0.13793) / 7.787;
+	else
+		*Z = Lab2XYZ->Z0 * tmp * tmp * tmp;
+}
+
 /* Process a buffer of data.
  */
 static void
@@ -97,33 +125,13 @@ vips_Lab2XYZ_line( VipsColour *colour, VipsPel *out, VipsPel **in, int width )
 	for( x = 0; x < width; x++ ) {
 		float L, a, b;
 		float X, Y, Z;
-		double cby, tmp;
 
 		L = p[0];
 		a = p[1];
 		b = p[2];
 		p += 3;
 
-		if( L < 8.0 ) {
-			Y = (L * Lab2XYZ->Y0) / 903.3;
-			cby = 7.787 * (Y / Lab2XYZ->Y0) + 16.0 / 116.0;
-		}
-		else {
-			cby = (L + 16.0) / 116.0;
-			Y = Lab2XYZ->Y0 * cby * cby * cby;
-		}
-
-		tmp = a / 500.0 + cby;
-		if( tmp < 0.2069 )
-			X = Lab2XYZ->X0 * (tmp - 0.13793) / 7.787;
-		else    
-			X = Lab2XYZ->X0 * tmp * tmp * tmp;
-
-		tmp = cby - b / 200.0;
-		if( tmp < 0.2069 )
-			Z = Lab2XYZ->Z0 * (tmp - 0.13793) / 7.787;
-		else    
-			Z = Lab2XYZ->Z0 * tmp * tmp * tmp;
+		vips_col_Lab2XYZ_helper( Lab2XYZ, L, a, b, &X, &Y, &Z );
 
 		/* Write.
 		 */
@@ -234,23 +242,11 @@ vips_Lab2XYZ( VipsImage *in, VipsImage **out, ... )
  */
 void
 vips_col_Lab2XYZ( float L, float a, float b, float *X, float *Y, float *Z )
-{	
-	float in[3];
-	float *x;
-	float out[3];
+{
 	VipsLab2XYZ Lab2XYZ;
 
-	in[0] = L;
-	in[1] = a;
-	in[2] = b;
-	x = in;
 	Lab2XYZ.X0 = VIPS_D65_X0;
 	Lab2XYZ.Y0 = VIPS_D65_Y0;
 	Lab2XYZ.Z0 = VIPS_D65_Z0;
-	vips_Lab2XYZ_line( (VipsColour *) &Lab2XYZ, 
-		(VipsPel *) out, (VipsPel **) &x, 1 );
-	*X = out[0];
-	*Y = out[1];
-	*Z = out[2];
+	vips_col_Lab2XYZ_helper( &Lab2XYZ, L, a, b, X, Y, Z );
 }
-
diff --git a/libvips/colour/XYZ2Lab.c b/libvips/colour/XYZ2Lab.c
index f7a959c0..4770364a 100644
--- a/libvips/colour/XYZ2Lab.c
+++ b/libvips/colour/XYZ2Lab.c
@@ -107,49 +107,67 @@ table_init( void *client )
 	return( NULL );
 }
 
+static void
+vips_col_XYZ2Lab_helper( VipsXYZ2Lab *XYZ2Lab,
+	float X, float Y, float Z, float *L, float *a, float *b )
+{
+	static GOnce once = G_ONCE_INIT;
+
+	VIPS_ONCE( &once, table_init, NULL );
+
+	float nX, nY, nZ;
+	int i;
+	float f;
+	float cbx, cby, cbz;
+
+	nX = QUANT_ELEMENTS * X / XYZ2Lab->X0;
+	nY = QUANT_ELEMENTS * Y / XYZ2Lab->Y0;
+	nZ = QUANT_ELEMENTS * Z / XYZ2Lab->Z0;
+
+	/* CLIP is much faster than FCLIP, and we want an int result.
+	 */
+	i = VIPS_CLIP( 0, (int) nX, QUANT_ELEMENTS - 2 );
+	f = nX - i;
+	cbx = cbrt_table[i] + f * (cbrt_table[i + 1] - cbrt_table[i]);
+
+	i = VIPS_CLIP( 0, (int) nY, QUANT_ELEMENTS - 2 );
+	f = nY - i;
+	cby = cbrt_table[i] + f * (cbrt_table[i + 1] - cbrt_table[i]);
+
+	i = VIPS_CLIP( 0, (int) nZ, QUANT_ELEMENTS - 2 );
+	f = nZ - i;
+	cbz = cbrt_table[i] + f * (cbrt_table[i + 1] - cbrt_table[i]);
+
+	*L = 116.0 * cby - 16.0;
+	*a = 500.0 * (cbx - cby);
+	*b = 200.0 * (cby - cbz);
+}
+
 /* Process a buffer of data.
  */
 static void
 vips_XYZ2Lab_line( VipsColour *colour, VipsPel *out, VipsPel **in, int width )
 {
-	static GOnce once = G_ONCE_INIT;
-
 	VipsXYZ2Lab *XYZ2Lab = (VipsXYZ2Lab *) colour;
 	float *p = (float *) in[0];
 	float *q = (float *) out;
 
 	int x;
 
-	VIPS_ONCE( &once, table_init, NULL );
-
 	for( x = 0; x < width; x++ ) {
-		float nX, nY, nZ;
-		int i;
-		float f;
-		float cbx, cby, cbz;
+		float X, Y, Z;
+		float L, a, b;
 
-		nX = QUANT_ELEMENTS * p[0] / XYZ2Lab->X0;
-		nY = QUANT_ELEMENTS * p[1] / XYZ2Lab->Y0;
-		nZ = QUANT_ELEMENTS * p[2] / XYZ2Lab->Z0;
+		X = p[0];
+		Y = p[1];
+		Z = p[2];
 		p += 3;
 
-		/* CLIP is much faster than FCLIP, and we want an int result.
-		 */
-		i = VIPS_CLIP( 0, (int) nX, QUANT_ELEMENTS - 2 );
-		f = nX - i;
-		cbx = cbrt_table[i] + f * (cbrt_table[i + 1] - cbrt_table[i]);
+		vips_col_XYZ2Lab_helper( XYZ2Lab, X, Y, Z, &L, &a, &b );
 
-		i = VIPS_CLIP( 0, (int) nY, QUANT_ELEMENTS - 2 );
-		f = nY - i;
-		cby = cbrt_table[i] + f * (cbrt_table[i + 1] - cbrt_table[i]);
-
-		i = VIPS_CLIP( 0, (int) nZ, QUANT_ELEMENTS - 2 );
-		f = nZ - i;
-		cbz = cbrt_table[i] + f * (cbrt_table[i + 1] - cbrt_table[i]);
-
-		q[0] = 116.0 * cby - 16.0;
-		q[1] = 500.0 * (cbx - cby);
-		q[2] = 200.0 * (cby - cbz);
+		q[0] = L;
+		q[1] = a;
+		q[2] = b;
 		q += 3;
 	}
 }
@@ -169,24 +187,13 @@ vips_XYZ2Lab_line( VipsColour *colour, VipsPel *out, VipsPel **in, int width )
  */
 void
 vips_col_XYZ2Lab( float X, float Y, float Z, float *L, float *a, float *b )
-{	
-	float in[3];
-	float out[3];
-	float *x;
+{
 	VipsXYZ2Lab XYZ2Lab;
 
-	in[0] = X;
-	in[1] = Y;
-	in[2] = Z;
-	x = in;
 	XYZ2Lab.X0 = VIPS_D65_X0;
 	XYZ2Lab.Y0 = VIPS_D65_Y0;
 	XYZ2Lab.Z0 = VIPS_D65_Z0;
-	vips_XYZ2Lab_line( (VipsColour *) &XYZ2Lab, 
-		(VipsPel *) out, (VipsPel **) &x, 1 );
-	*L = out[0];
-	*a = out[1];
-	*b = out[2];
+	vips_col_XYZ2Lab_helper( &XYZ2Lab, X, Y, Z, L, a, b );
 }
 
 static int
@@ -273,4 +280,3 @@ vips_XYZ2Lab( VipsImage *in, VipsImage **out, ... )
 
 	return( result );
 }
-
diff --git a/libvips/foreign/jp2ksave.c b/libvips/foreign/jp2ksave.c
index 62420b1b..bb749ac9 100644
--- a/libvips/foreign/jp2ksave.c
+++ b/libvips/foreign/jp2ksave.c
@@ -828,16 +828,16 @@ vips_foreign_save_jp2k_build( VipsObject *object )
 	jp2k->parameters.cp_tdx = jp2k->tile_width;
 	jp2k->parameters.cp_tdy = jp2k->tile_height;
 
-	/* Makes three band images smaller, somehow.
+	/* Makes many-band, non-subsampled images smaller, somehow.
 	 */
-	jp2k->parameters.tcp_mct = save->ready->Bands >= 3 ? 1 : 0;
+	jp2k->parameters.tcp_mct = save->ready->Bands >= 3 && !jp2k->subsample;
 
 	/* Number of layers to write. Smallest layer is c. 2^5 on the smallest
 	 * axis.
 	 */
 	jp2k->parameters.numresolution = VIPS_MAX( 1, 
 		log( VIPS_MIN( save->ready->Xsize, save->ready->Ysize ) ) / 
-		log( 2 ) - 4 );
+		log( 2 ) - 5 );
 #ifdef DEBUG
 	printf( "vips_foreign_save_jp2k_build: numresolutions = %d\n", 
 		jp2k->parameters.numresolution );