diff --git a/example/font_outline.cpp b/example/font_outline.cpp
index 9535d53..14269fd 100644
--- a/example/font_outline.cpp
+++ b/example/font_outline.cpp
@@ -59,15 +59,15 @@ uint32_t transform(ta_parameter_writer& parameter,
 
     auto polygon = global_polygon_type_0(texture_address);
     polygon.parameter_control_word = para_control::para_type::polygon_or_modifier_volume
-      | para_control::list_type::opaque
-      | obj_control::col_type::packed_color
-      | obj_control::texture;
+				   | para_control::list_type::opaque
+				   | obj_control::col_type::packed_color
+				   | obj_control::texture;
 
     polygon.tsp_instruction_word = tsp_instruction_word::src_alpha_instr::one
-      | tsp_instruction_word::dst_alpha_instr::zero
-      | tsp_instruction_word::fog_control::no_fog
-      | tsp_instruction_word::texture_u_size::from_int(texture_width)
-      | tsp_instruction_word::texture_v_size::from_int(texture_height);
+				 | tsp_instruction_word::dst_alpha_instr::zero
+				 | tsp_instruction_word::fog_control::no_fog
+				 | tsp_instruction_word::texture_u_size::from_int(texture_width)
+				 | tsp_instruction_word::texture_v_size::from_int(texture_height);
 
     polygon.texture_control_word = texture_control_word::pixel_format::_8bpp_palette
 				 | texture_control_word::scan_order::twiddled
@@ -266,10 +266,11 @@ void main()
 
     auto parameter = ta_parameter_writer(ta_parameter_buf);
 
+    /*
     transform2(parameter,
 	       font->texture_width, font->texture_height);
+    */
 
-    /*
     transform(parameter,
 	      font->texture_width, font->texture_height,
 	      font->first_char_code,
@@ -283,7 +284,6 @@ void main()
 	      glyphs,
 	      cabal, 26,
 	      font->glyph_height * 1);
-    */
 
     parameter.append<global_end_of_list>() = global_end_of_list();
 
diff --git a/twiddle.hpp b/twiddle.hpp
index 61a960e..7c521aa 100644
--- a/twiddle.hpp
+++ b/twiddle.hpp
@@ -183,17 +183,21 @@ void texture3(volatile T * dst, const U * src,
   constexpr uint32_t pixels_per_u = u_bits / src_bits_per_pixel;
   static_assert(pixels_per_u == 1 || pixels_per_u == 2 || pixels_per_u == 4 || pixels_per_u == 8 || pixels_per_u == 16 || pixels_per_u == 32);
 
+  constexpr uint32_t src_val_mask = ((1 << src_bits_per_pixel) - 1);
+  constexpr uint32_t dst_val_mask = ((1 << dst_bits_per_pixel) - 1);
+  constexpr uint32_t dst_src_shift = (dst_bits_per_pixel < src_bits_per_pixel) ? (src_bits_per_pixel - dst_bits_per_pixel) : 0;
+
   T dst_val = 0;
   for (uint32_t curve_ix = 0; curve_ix <= curve_end_ix; curve_ix++) {
     auto [x, y] = from_ix(curve_ix);
     const uint32_t src_ix = y * src_stride + (x / pixels_per_u);
     const uint32_t src_ix_mod = x & (pixels_per_u - 1);
-    const U src_val = (src[src_ix] >> (src_bits_per_pixel * src_ix_mod)) & ((1 << src_bits_per_pixel) - 1);
+    const U src_val = (src[src_ix] >> (src_bits_per_pixel * src_ix_mod)) & src_val_mask;
     if constexpr (pixels_per_t == 1) {
       dst[curve_ix] = src_val;
     } else {
       const uint32_t curve_ix_mod = curve_ix & (pixels_per_t - 1);
-      dst_val |= src_val << (dst_bits_per_pixel * curve_ix_mod);
+      dst_val |= ((src_val >> dst_src_shift) & dst_val_mask) << (dst_bits_per_pixel * curve_ix_mod);
 
       if (curve_ix_mod == (pixels_per_t - 1)) {
 	dst[curve_ix / pixels_per_t] = dst_val;