diff --git a/src/dec/buffer.c b/src/dec/buffer.c
index 13e01ee0..ad953737 100644
--- a/src/dec/buffer.c
+++ b/src/dec/buffer.c
@@ -22,7 +22,10 @@ extern "C" {
 // WebPDecBuffer
 
 // Number of bytes per pixel for the different color-spaces.
-static const int kModeBpp[MODE_LAST] = { 3, 4, 3, 4, 4, 2, 2, 1, 1 };
+static const int kModeBpp[MODE_LAST] = {
+  3, 4, 3, 4, 4, 2, 2,
+  4, 4, 4, 2,    // pre-multiplied modes
+  1, 1 };
 
 // Check that webp_csp_mode is within the bounds of WEBP_CSP_MODE.
 // Convert to an integer to handle both the unsigned/signed enum cases
@@ -38,7 +41,7 @@ static VP8StatusCode CheckDecBuffer(const WebPDecBuffer* const buffer) {
   const int height = buffer->height;
   if (!IsValidColorspace(mode)) {
     ok = 0;
-  } else if (mode >= MODE_YUV) {   // YUV checks
+  } else if (!WebPIsRGBMode(mode)) {   // YUV checks
     const WebPYUVABuffer* const buf = &buffer->u.YUVA;
     const size_t size = buf->y_stride * height;
     const size_t u_size = buf->u_stride * ((height + 1) / 2);
@@ -81,7 +84,7 @@ static VP8StatusCode AllocateBuffer(WebPDecBuffer* const buffer) {
     const int stride = w * kModeBpp[mode];
     const uint64_t size = (uint64_t)stride * h;
 
-    if (mode >= MODE_YUV) {
+    if (!WebPIsRGBMode(mode)) {
       uv_stride = (w + 1) / 2;
       uv_size = (uint64_t)uv_stride * ((h + 1) / 2);
       if (mode == MODE_YUVA) {
@@ -101,7 +104,7 @@ static VP8StatusCode AllocateBuffer(WebPDecBuffer* const buffer) {
       return VP8_STATUS_OUT_OF_MEMORY;
     }
 
-    if (mode >= MODE_YUV) {   // YUVA initialization
+    if (!WebPIsRGBMode(mode)) {   // YUVA initialization
       WebPYUVABuffer* const buf = &buffer->u.YUVA;
       buf->y = output;
       buf->y_stride = stride;
diff --git a/src/dec/frame.c b/src/dec/frame.c
index 50982dbd..16666627 100644
--- a/src/dec/frame.c
+++ b/src/dec/frame.c
@@ -200,7 +200,12 @@ static int FinishRow(VP8Decoder* const dec, VP8Io* const io) {
       y_end = io->crop_bottom;    // make sure we don't overflow on last row.
     }
     io->a = NULL;
-    if (dec->alpha_data_ && y_start < y_end) {
+    if (dec->alpha_data_ != NULL && y_start < y_end) {
+      // TODO(skal): several things to correct here:
+      // * testing presence of alpha with dec->alpha_data_ is not a good idea
+      // * we're actually decompressing the full plane only once. It should be
+      //   more obvious from signature.
+      // * we could free alpha_data_ right after this call, but we don't own.
       io->a = VP8DecompressAlphaRows(dec, y_start, y_end - y_start);
       if (io->a == NULL) {
         return VP8SetError(dec, VP8_STATUS_BITSTREAM_ERROR,
@@ -214,7 +219,7 @@ static int FinishRow(VP8Decoder* const dec, VP8Io* const io) {
       io->y += dec->cache_y_stride_ * delta_y;
       io->u += dec->cache_uv_stride_ * (delta_y >> 1);
       io->v += dec->cache_uv_stride_ * (delta_y >> 1);
-      if (io->a) {
+      if (io->a != NULL) {
         io->a += io->width * delta_y;
       }
     }
@@ -222,7 +227,7 @@ static int FinishRow(VP8Decoder* const dec, VP8Io* const io) {
       io->y += io->crop_left;
       io->u += io->crop_left >> 1;
       io->v += io->crop_left >> 1;
-      if (io->a) {
+      if (io->a != NULL) {
         io->a += io->crop_left;
       }
       io->mb_y = y_start - io->crop_top;
@@ -417,7 +422,8 @@ static int AllocateMemory(VP8Decoder* const dec) {
                             + kFilterExtraRows[dec->filter_type_]) * 3 / 2;
   const size_t cache_size = top_size * cache_height;
   const size_t alpha_size =
-      dec->alpha_data_ ? (dec->pic_hdr_.width_ * dec->pic_hdr_.height_) : 0;
+      (dec->alpha_data_ != NULL) ? dec->pic_hdr_.width_ * dec->pic_hdr_.height_
+                                 : 0;
   const size_t needed = intra_pred_mode_size
                       + top_size + mb_info_size + f_info_size
                       + yuv_size + coeffs_size
diff --git a/src/dec/io.c b/src/dec/io.c
index 2760a29d..82c7c0d3 100644
--- a/src/dec/io.c
+++ b/src/dec/io.c
@@ -52,9 +52,7 @@ static int EmitSampledRGB(const VP8Io* const io, WebPDecParams* const p) {
   const uint8_t* y_src = io->y;
   const uint8_t* u_src = io->u;
   const uint8_t* v_src = io->v;
-  const WebPSampleLinePairFunc sample =
-      io->a ? WebPSamplersKeepAlpha[output->colorspace]
-            : WebPSamplers[output->colorspace];
+  const WebPSampleLinePairFunc sample = WebPSamplers[output->colorspace];
   const int mb_w = io->mb_w;
   const int last = io->mb_h - 1;
   int j;
@@ -106,9 +104,7 @@ static int EmitFancyRGB(const VP8Io* const io, WebPDecParams* const p) {
   int num_lines_out = io->mb_h;   // a priori guess
   const WebPRGBABuffer* const buf = &p->output->u.RGBA;
   uint8_t* dst = buf->rgba + io->mb_y * buf->stride;
-  const WebPUpsampleLinePairFunc upsample =
-      io->a ? WebPUpsamplersKeepAlpha[p->output->colorspace]
-            : WebPUpsamplers[p->output->colorspace];
+  WebPUpsampleLinePairFunc upsample = WebPUpsamplers[p->output->colorspace];
   const uint8_t* cur_y = io->y;
   const uint8_t* cur_u = io->u;
   const uint8_t* cur_v = io->v;
@@ -124,8 +120,6 @@ static int EmitFancyRGB(const VP8Io* const io, WebPDecParams* const p) {
     upsample(NULL, cur_y, cur_u, cur_v, cur_u, cur_v, NULL, dst, mb_w);
   } else {
     // We can finish the left-over line from previous call.
-    // Warning! Don't overwrite the alpha values (if any), as they
-    // are not lagging one line behind but are already written.
     upsample(p->tmp_y, cur_y, top_u, top_v, cur_u, cur_v,
              dst - buf->stride, dst, mb_w);
     ++num_lines_out;
@@ -184,45 +178,73 @@ static int EmitAlphaYUV(const VP8Io* const io, WebPDecParams* const p) {
 }
 
 static int EmitAlphaRGB(const VP8Io* const io, WebPDecParams* const p) {
-  const int mb_w = io->mb_w;
-  const int mb_h = io->mb_h;
-  int i, j;
-  const WebPRGBABuffer* const buf = &p->output->u.RGBA;
-  uint8_t* dst = buf->rgba + io->mb_y * buf->stride +
-                 (p->output->colorspace == MODE_ARGB ? 0 : 3);
   const uint8_t* alpha = io->a;
-  if (alpha) {
-    for (j = 0; j < mb_h; ++j) {
-      for (i = 0; i < mb_w; ++i) {
-        dst[4 * i] = alpha[i];
+  if (alpha != NULL) {
+    const int mb_w = io->mb_w;
+    const int mb_h = io->mb_h;
+    int i, j;
+    const WEBP_CSP_MODE colorspace = p->output->colorspace;
+    const int alpha_first =
+        (colorspace == MODE_ARGB || colorspace == MODE_Argb);
+    const WebPRGBABuffer* const buf = &p->output->u.RGBA;
+    int start_y = io->mb_y;
+    int num_rows = mb_h;
+
+    // We compensate for the 1-line delay of fancy upscaler.
+    // This is similar to EmitFancyRGB().
+    if (io->fancy_upsampling) {
+      if (start_y == 0) {
+        // We don't process the last row yet. It'll be done during next call.
+        --num_rows;
+      } else {
+        --start_y;
+        // Fortunately, *alpha data is persistent, so we can go back
+        // one row and finish alpha blending, now that the fancy upscaler
+        // completed the YUV->RGB interpolation.
+        alpha -= io->width;
+      }
+      if (io->crop_top + io->mb_y + mb_h == io->crop_bottom) {
+        // If it's the very last call, we process all the remaing rows!
+        num_rows = io->crop_bottom - io->crop_top - start_y;
+      }
+    }
+    {
+      uint8_t* const base_rgba = buf->rgba + start_y * buf->stride;
+      uint8_t* dst = base_rgba + (alpha_first ? 0 : 3);
+      for (j = 0; j < num_rows; ++j) {
+        for (i = 0; i < mb_w; ++i) dst[4 * i] = alpha[i];
+        alpha += io->width;
+        dst += buf->stride;
+      }
+      if (WebPIsPremultipliedMode(colorspace)) {
+        WebPApplyAlphaMultiply(base_rgba, alpha_first,
+                               mb_w, num_rows, buf->stride);
       }
-      alpha += io->width;
-      dst += buf->stride;
     }
   }
   return 0;
 }
 
-static WEBP_INLINE uint32_t clip(uint32_t v, uint32_t max_value) {
-  return (v > max_value) ? max_value : v;
-}
-
 static int EmitAlphaRGBA4444(const VP8Io* const io, WebPDecParams* const p) {
-  const int mb_w = io->mb_w;
-  const int mb_h = io->mb_h;
-  int i, j;
-  const WebPRGBABuffer* const buf = &p->output->u.RGBA;
-  uint8_t* dst = buf->rgba + io->mb_y * buf->stride + 1;
   const uint8_t* alpha = io->a;
-  if (alpha) {
+  if (alpha != NULL) {
+    const int mb_w = io->mb_w;
+    const int mb_h = io->mb_h;
+    int i, j;
+    const WebPRGBABuffer* const buf = &p->output->u.RGBA;
+    uint8_t* const base_rgba = buf->rgba + io->mb_y * buf->stride;
+    uint8_t* alpha_dst = base_rgba + 1;
     for (j = 0; j < mb_h; ++j) {
       for (i = 0; i < mb_w; ++i) {
         // Fill in the alpha value (converted to 4 bits).
-        const uint32_t alpha_val = clip((alpha[i] + 8) >> 4, 15);
-        dst[2 * i] = (dst[2 * i] & 0xf0) | alpha_val;
+        const uint32_t alpha_val = VP8Clip4Bits(alpha[i]);
+        alpha_dst[2 * i] = (alpha_dst[2 * i] & 0xf0) | alpha_val;
       }
       alpha += io->width;
-      dst += buf->stride;
+      alpha_dst += buf->stride;
+    }
+    if (p->output->colorspace == MODE_rgbA_4444) {
+      WebPApplyAlphaMultiply4444(base_rgba, mb_w, mb_h, buf->stride);
     }
   }
   return 0;
@@ -259,13 +281,8 @@ static int EmitRescaledAlphaYUV(const VP8Io* const io, WebPDecParams* const p) {
   return 0;
 }
 
-static int IsAlphaMode(WEBP_CSP_MODE mode) {
-  return (mode == MODE_RGBA || mode == MODE_BGRA || mode == MODE_ARGB ||
-          mode == MODE_RGBA_4444 || mode == MODE_YUVA);
-}
-
 static int InitYUVRescaler(const VP8Io* const io, WebPDecParams* const p) {
-  const int has_alpha = IsAlphaMode(p->output->colorspace);
+  const int has_alpha = WebPIsAlphaMode(p->output->colorspace);
   const WebPYUVABuffer* const buf = &p->output->u.YUVA;
   const int out_width  = io->scaled_width;
   const int out_height = io->scaled_height;
@@ -302,6 +319,7 @@ static int InitYUVRescaler(const VP8Io* const io, WebPDecParams* const p) {
                    uv_in_height, uv_out_height,
                    work + work_size + uv_work_size);
   p->emit = EmitRescaledYUV;
+
   if (has_alpha) {
     WebPRescalerInit(&p->scaler_a, io->mb_w, io->mb_h,
                      buf->a, out_width, out_height, buf->a_stride, 1,
@@ -364,59 +382,73 @@ static int EmitRescaledRGB(const VP8Io* const io, WebPDecParams* const p) {
 
 static int ExportAlpha(WebPDecParams* const p, int y_pos) {
   const WebPRGBABuffer* const buf = &p->output->u.RGBA;
-  uint8_t* dst = buf->rgba + (p->last_y + y_pos) * buf->stride +
-                 (p->output->colorspace == MODE_ARGB ? 0 : 3);
+  uint8_t* const base_rgba = buf->rgba + (p->last_y + y_pos) * buf->stride;
+  const WEBP_CSP_MODE colorspace = p->output->colorspace;
+  const int alpha_first =
+      (colorspace == MODE_ARGB || colorspace == MODE_Argb);
+  uint8_t* dst = base_rgba + (alpha_first ? 0 : 3);
   int num_lines_out = 0;
+  const int is_premult_alpha = WebPIsPremultipliedMode(colorspace);
+  const int width = p->scaler_a.dst_width;
+
   while (WebPRescalerHasPendingOutput(&p->scaler_a)) {
     int i;
     assert(p->last_y + y_pos + num_lines_out < p->output->height);
     WebPRescalerExportRow(&p->scaler_a);
-    for (i = 0; i < p->scaler_a.dst_width; ++i) {
-      dst[4 * i] = p->scaler_a.dst[i];
-    }
+    for (i = 0; i < width; ++i) dst[4 * i] = p->scaler_a.dst[i];
     dst += buf->stride;
     ++num_lines_out;
   }
+  if (is_premult_alpha) {
+    WebPApplyAlphaMultiply(base_rgba, alpha_first,
+                           width, num_lines_out, buf->stride);
+  }
   return num_lines_out;
 }
 
 static int ExportAlphaRGBA4444(WebPDecParams* const p, int y_pos) {
   const WebPRGBABuffer* const buf = &p->output->u.RGBA;
-  uint8_t* dst = buf->rgba + (p->last_y + y_pos) * buf->stride + 1;
+  uint8_t* base_rgba = buf->rgba + (p->last_y + y_pos) * buf->stride;
+  uint8_t* alpha_dst = base_rgba + 1;
   int num_lines_out = 0;
+  const WEBP_CSP_MODE colorspace = p->output->colorspace;
+  const int width = p->scaler_a.dst_width;
+  const int is_premult_alpha = WebPIsPremultipliedMode(colorspace);
+
   while (WebPRescalerHasPendingOutput(&p->scaler_a)) {
     int i;
     assert(p->last_y + y_pos + num_lines_out < p->output->height);
     WebPRescalerExportRow(&p->scaler_a);
-    for (i = 0; i < p->scaler_a.dst_width; ++i) {
+    for (i = 0; i < width; ++i) {
       // Fill in the alpha value (converted to 4 bits).
-      const uint32_t alpha_val = clip((p->scaler_a.dst[i] + 8) >> 4, 15);
-      dst[2 * i] = (dst[2 * i] & 0xf0) | alpha_val;
+      const uint32_t alpha_val = VP8Clip4Bits(p->scaler_a.dst[i]);
+      alpha_dst[2 * i] = (alpha_dst[2 * i] & 0xf0) | alpha_val;
     }
-    dst += buf->stride;
+    alpha_dst += buf->stride;
     ++num_lines_out;
   }
+  if (is_premult_alpha) {
+    WebPApplyAlphaMultiply4444(base_rgba, width, num_lines_out, buf->stride);
+  }
   return num_lines_out;
 }
 
 static int EmitRescaledAlphaRGB(const VP8Io* const io, WebPDecParams* const p) {
   if (io->a != NULL) {
-    int (* const output_func)(WebPDecParams* const, int) =
-        (p->output->colorspace == MODE_RGBA_4444) ? ExportAlphaRGBA4444
-                                                  : ExportAlpha;
     WebPRescaler* const scaler = &p->scaler_a;
-    int j = 0, pos = 0;
+    int j = 0;
+    int pos = 0;
     while (j < io->mb_h) {
       j += WebPRescalerImport(scaler, io->mb_h - j,
                               io->a + j * io->width, io->width);
-      pos += output_func(p, pos);
+      pos += p->emit_alpha_row(p, pos);
     }
   }
   return 0;
 }
 
 static int InitRGBRescaler(const VP8Io* const io, WebPDecParams* const p) {
-  const int has_alpha = IsAlphaMode(p->output->colorspace);
+  const int has_alpha = WebPIsAlphaMode(p->output->colorspace);
   const int out_width  = io->scaled_width;
   const int out_height = io->scaled_height;
   const int uv_in_width  = (io->mb_w + 1) >> 1;
@@ -459,6 +491,12 @@ static int InitRGBRescaler(const VP8Io* const io, WebPDecParams* const p) {
                      io->mb_w, out_width, io->mb_h, out_height,
                      work + 3 * work_size);
     p->emit_alpha = EmitRescaledAlphaRGB;
+    if (p->output->colorspace == MODE_RGBA_4444 ||
+        p->output->colorspace == MODE_rgbA_4444) {
+      p->emit_alpha_row = ExportAlphaRGBA4444;
+    } else {
+      p->emit_alpha_row = ExportAlpha;
+    }
   }
   return 1;
 }
@@ -469,12 +507,13 @@ static int InitRGBRescaler(const VP8Io* const io, WebPDecParams* const p) {
 static int CustomSetup(VP8Io* io) {
   WebPDecParams* const p = (WebPDecParams*)io->opaque;
   const WEBP_CSP_MODE colorspace = p->output->colorspace;
-  const int is_rgb = (colorspace < MODE_YUV);
-  const int is_alpha = IsAlphaMode(colorspace);
+  const int is_rgb = WebPIsRGBMode(colorspace);
+  const int is_alpha = WebPIsAlphaMode(colorspace);
 
   p->memory = NULL;
   p->emit = NULL;
   p->emit_alpha = NULL;
+  p->emit_alpha_row = NULL;
   if (!WebPIoInitFromOptions(p->options, io, is_alpha ? MODE_YUV : MODE_YUVA)) {
     return 0;
   }
@@ -505,10 +544,12 @@ static int CustomSetup(VP8Io* io) {
       p->emit = EmitYUV;
     }
     if (is_alpha) {  // need transparency output
+      if (WebPIsPremultipliedMode(colorspace)) WebPInitPremultiply();
       p->emit_alpha =
-          is_rgb ? (colorspace == MODE_RGBA_4444 ? EmitAlphaRGBA4444
-                                                 : EmitAlphaRGB)
-                 : EmitAlphaYUV;
+          (colorspace == MODE_RGBA_4444 || colorspace == MODE_rgbA_4444) ?
+              EmitAlphaRGBA4444
+          : is_rgb ? EmitAlphaRGB
+          : EmitAlphaYUV;
     }
   }
 
diff --git a/src/dec/vp8i.h b/src/dec/vp8i.h
index d811ca51..af96dc91 100644
--- a/src/dec/vp8i.h
+++ b/src/dec/vp8i.h
@@ -276,7 +276,7 @@ struct VP8Decoder {
   // extensions
   const uint8_t* alpha_data_;   // compressed alpha data (if present)
   size_t alpha_data_size_;
-  uint8_t* alpha_plane_;        // output
+  uint8_t* alpha_plane_;        // output. Persistent, contains the whole data.
 
   int layer_colorspace_;
   const uint8_t* layer_data_;   // compressed layer data (if present)
diff --git a/src/dec/webp.c b/src/dec/webp.c
index 22e998b4..67614796 100644
--- a/src/dec/webp.c
+++ b/src/dec/webp.c
@@ -477,7 +477,7 @@ static uint8_t* Decode(WEBP_CSP_MODE mode, const uint8_t* data,
     WebPCopyDecBuffer(&output, keep_info);
   }
   // return decoded samples (don't clear 'output'!)
-  return (mode >= MODE_YUV) ? output.u.YUVA.y : output.u.RGBA.rgba;
+  return WebPIsRGBMode(mode) ? output.u.RGBA.rgba : output.u.YUVA.y;
 }
 
 uint8_t* WebPDecodeRGB(const uint8_t* data, size_t data_size,
@@ -685,7 +685,7 @@ int WebPIoInitFromOptions(const WebPDecoderOptions* const options,
     h = options->crop_height;
     x = options->crop_left;
     y = options->crop_top;
-    if (src_colorspace >= MODE_YUV) {   // only snap for YUV420 or YUV422
+    if (!WebPIsRGBMode(src_colorspace)) {   // only snap for YUV420 or YUV422
       x &= ~1;
       y &= ~1;    // TODO(later): only for YUV420, not YUV422.
     }
diff --git a/src/dec/webpi.h b/src/dec/webpi.h
index 9d29e633..527a3888 100644
--- a/src/dec/webpi.h
+++ b/src/dec/webpi.h
@@ -24,6 +24,7 @@ extern "C" {
 
 typedef struct WebPDecParams WebPDecParams;
 typedef int (*OutputFunc)(const VP8Io* const io, WebPDecParams* const p);
+typedef int (*OutputRowFunc)(WebPDecParams* const p, int y_pos);
 
 struct WebPDecParams {
   WebPDecBuffer* output;             // output buffer.
@@ -34,9 +35,11 @@ struct WebPDecParams {
   const WebPDecoderOptions* options;  // if not NULL, use alt decoding features
   // rescalers
   WebPRescaler scaler_y, scaler_u, scaler_v, scaler_a;
-  void* memory;               // overall scratch memory for the output work.
-  OutputFunc emit;            // output RGB or YUV samples
-  OutputFunc emit_alpha;      // output alpha channel
+  void* memory;                  // overall scratch memory for the output work.
+
+  OutputFunc emit;               // output RGB or YUV samples
+  OutputFunc emit_alpha;         // output alpha channel
+  OutputRowFunc emit_alpha_row;  // output one line of rescaled alpha values
 };
 
 // Should be called first, before any use of the WebPDecParams object.
diff --git a/src/dsp/dsp.h b/src/dsp/dsp.h
index db9c1455..ff78b4c7 100644
--- a/src/dsp/dsp.h
+++ b/src/dsp/dsp.h
@@ -155,7 +155,6 @@ typedef void (*WebPUpsampleLinePairFunc)(
 
 // Fancy upsampling functions to convert YUV to RGB(A) modes
 extern WebPUpsampleLinePairFunc WebPUpsamplers[/* MODE_LAST */];
-extern WebPUpsampleLinePairFunc WebPUpsamplersKeepAlpha[/* MODE_LAST */];
 
 // Initializes SSE2 version of the fancy upsamplers.
 void WebPInitUpsamplersSSE2(void);
@@ -169,7 +168,6 @@ typedef void (*WebPSampleLinePairFunc)(
     uint8_t* top_dst, uint8_t* bottom_dst, int len);
 
 extern const WebPSampleLinePairFunc WebPSamplers[/* MODE_LAST */];
-extern const WebPSampleLinePairFunc WebPSamplersKeepAlpha[/* MODE_LAST */];
 
 // YUV444->RGB converters
 typedef void (*WebPYUV444Converter)(const uint8_t* y,
@@ -181,6 +179,23 @@ extern const WebPYUV444Converter WebPYUV444Converters[/* MODE_LAST */];
 // Main function to be called
 void WebPInitUpsamplers(void);
 
+//------------------------------------------------------------------------------
+// Pre-multiply planes with alpha values
+
+// Apply alpha pre-multiply on an rgba, bgra or argb plane of size w * h.
+// alpha_first should be 0 for argb, 1 for rgba or bgra (where alpha is last).
+extern void (*WebPApplyAlphaMultiply)(
+    uint8_t* rgba, int alpha_first, int w, int h, int stride);
+
+// Same, buf specifically for RGBA4444 format
+extern void (*WebPApplyAlphaMultiply4444)(
+    uint8_t* rgba4444, int w, int h, int stride);
+
+// To be called first before using the above.
+void WebPInitPremultiply(void);
+
+void WebPInitPremultiplySSE2(void);   // should not be called directly.
+
 //------------------------------------------------------------------------------
 
 #if defined(__cplusplus) || defined(c_plusplus)
diff --git a/src/dsp/lossless.c b/src/dsp/lossless.c
index 51745875..e5a51ca7 100644
--- a/src/dsp/lossless.c
+++ b/src/dsp/lossless.c
@@ -19,6 +19,8 @@ extern "C" {
 #include <stdlib.h>
 #include "./lossless.h"
 #include "../dec/vp8li.h"
+#include "../dsp/yuv.h"
+#include "../dsp/dsp.h"
 
 #ifdef USE_LOSSLESS_ENCODER
 
@@ -1041,8 +1043,7 @@ static void CopyOrSwap(const uint32_t* src, int num_pixels, uint8_t* dst,
 }
 
 void VP8LConvertFromBGRA(const uint32_t* const in_data, int num_pixels,
-                        WEBP_CSP_MODE out_colorspace,
-                        uint8_t* const rgba) {
+                         WEBP_CSP_MODE out_colorspace, uint8_t* const rgba) {
   switch (out_colorspace) {
     case MODE_RGB:
       ConvertBGRAToRGB(in_data, num_pixels, rgba);
@@ -1050,18 +1051,34 @@ void VP8LConvertFromBGRA(const uint32_t* const in_data, int num_pixels,
     case MODE_RGBA:
       ConvertBGRAToRGBA(in_data, num_pixels, rgba);
       break;
+    case MODE_rgbA:
+      ConvertBGRAToRGBA(in_data, num_pixels, rgba);
+      WebPApplyAlphaMultiply(rgba, 0, num_pixels, 1, 0);
+      break;
     case MODE_BGR:
       ConvertBGRAToBGR(in_data, num_pixels, rgba);
       break;
     case MODE_BGRA:
       CopyOrSwap(in_data, num_pixels, rgba, 1);
       break;
+    case MODE_bgrA:
+      CopyOrSwap(in_data, num_pixels, rgba, 1);
+      WebPApplyAlphaMultiply(rgba, 0, num_pixels, 1, 0);
+      break;
     case MODE_ARGB:
       CopyOrSwap(in_data, num_pixels, rgba, 0);
       break;
+    case MODE_Argb:
+      CopyOrSwap(in_data, num_pixels, rgba, 0);
+      WebPApplyAlphaMultiply(rgba, 1, num_pixels, 1, 0);
+      break;
     case MODE_RGBA_4444:
       ConvertBGRAToRGBA4444(in_data, num_pixels, rgba);
       break;
+    case MODE_rgbA_4444:
+      ConvertBGRAToRGBA4444(in_data, num_pixels, rgba);
+      WebPApplyAlphaMultiply4444(rgba, num_pixels, 1, 0);
+      break;
     case MODE_RGB_565:
       ConvertBGRAToRGB565(in_data, num_pixels, rgba);
       break;
diff --git a/src/dsp/upsampling.c b/src/dsp/upsampling.c
index d5fd902c..094c7311 100644
--- a/src/dsp/upsampling.c
+++ b/src/dsp/upsampling.c
@@ -23,7 +23,6 @@ extern "C" {
 
 // Fancy upsampling functions to convert YUV to RGB
 WebPUpsampleLinePairFunc WebPUpsamplers[MODE_LAST];
-WebPUpsampleLinePairFunc WebPUpsamplersKeepAlpha[MODE_LAST];
 
 // Given samples laid out in a square as:
 //  [a b]
@@ -100,11 +99,6 @@ UPSAMPLE_FUNC(UpsampleBgraLinePair, VP8YuvToBgra, 4)
 UPSAMPLE_FUNC(UpsampleArgbLinePair, VP8YuvToArgb, 4)
 UPSAMPLE_FUNC(UpsampleRgba4444LinePair, VP8YuvToRgba4444, 2)
 UPSAMPLE_FUNC(UpsampleRgb565LinePair,  VP8YuvToRgb565,  2)
-// These variants don't erase the alpha value
-UPSAMPLE_FUNC(UpsampleRgbaKeepAlphaLinePair, VP8YuvToRgb, 4)
-UPSAMPLE_FUNC(UpsampleBgraKeepAlphaLinePair, VP8YuvToBgr, 4)
-UPSAMPLE_FUNC(UpsampleArgbKeepAlphaLinePair, VP8YuvToArgbKeepA, 4)
-UPSAMPLE_FUNC(UpsampleRgba4444KeepAlphaLinePair, VP8YuvToRgba4444KeepA, 2)
 
 #undef LOAD_UV
 #undef UPSAMPLE_FUNC
@@ -145,11 +139,6 @@ SAMPLE_FUNC(SampleBgraLinePair,     VP8YuvToBgra, 4)
 SAMPLE_FUNC(SampleArgbLinePair,     VP8YuvToArgb, 4)
 SAMPLE_FUNC(SampleRgba4444LinePair, VP8YuvToRgba4444, 2)
 SAMPLE_FUNC(SampleRgb565LinePair,   VP8YuvToRgb565, 2)
-// These variants don't erase the alpha value
-SAMPLE_FUNC(SampleRgbaKeepAlphaLinePair, VP8YuvToRgb, 4)
-SAMPLE_FUNC(SampleBgraKeepAlphaLinePair, VP8YuvToBgr, 4)
-SAMPLE_FUNC(SampleArgbKeepAlphaLinePair, VP8YuvToArgbKeepA, 4)
-SAMPLE_FUNC(SampleRgba4444KeepAlphaLinePair, VP8YuvToRgba4444KeepA, 2)
 
 #undef SAMPLE_FUNC
 
@@ -160,17 +149,11 @@ const WebPSampleLinePairFunc WebPSamplers[MODE_LAST] = {
   SampleBgraLinePair,      // MODE_BGRA
   SampleArgbLinePair,      // MODE_ARGB
   SampleRgba4444LinePair,  // MODE_RGBA_4444
-  SampleRgb565LinePair     // MODE_RGB_565
-};
-
-const WebPSampleLinePairFunc WebPSamplersKeepAlpha[MODE_LAST] = {
-  SampleRgbLinePair,                // MODE_RGB
-  SampleRgbaKeepAlphaLinePair,      // MODE_RGBA
-  SampleBgrLinePair,                // MODE_BGR
-  SampleBgraKeepAlphaLinePair,      // MODE_BGRA
-  SampleArgbKeepAlphaLinePair,      // MODE_ARGB
-  SampleRgba4444KeepAlphaLinePair,  // MODE_RGBA_4444
-  SampleRgb565LinePair              // MODE_RGB_565
+  SampleRgb565LinePair,    // MODE_RGB_565
+  SampleRgbaLinePair,      // MODE_rgbA
+  SampleBgraLinePair,      // MODE_bgrA
+  SampleArgbLinePair,      // MODE_Argb
+  SampleRgba4444LinePair   // MODE_rgbA_4444
 };
 
 //------------------------------------------------------------------------------
@@ -200,9 +183,82 @@ const WebPYUV444Converter WebPYUV444Converters[MODE_LAST] = {
   Yuv444ToBgra,      // MODE_BGRA
   Yuv444ToArgb,      // MODE_ARGB
   Yuv444ToRgba4444,  // MODE_RGBA_4444
-  Yuv444ToRgb565     // MODE_RGB_565
+  Yuv444ToRgb565,    // MODE_RGB_565
+  Yuv444ToRgba,      // MODE_rgbA
+  Yuv444ToBgra,      // MODE_bgrA
+  Yuv444ToArgb,      // MODE_Argb
+  Yuv444ToRgba4444   // MODE_rgbA_4444
 };
 
+//------------------------------------------------------------------------------
+// Premultiplied modes
+
+// non dithered-modes
+
+// (x * a * 32897) >> 23 is bit-wise equivalent to (int)(x * a / 255.)
+// for all 8bit x or a. For bit-wise equivalence to (int)(x * a / 255. + .5),
+// one can use instead: (x * a * 65793 + (1 << 23)) >> 24
+#if 1     // (int)(x * a / 255.)
+#define MULTIPLIER(a)   ((a) * 32897UL)
+#define PREMULTIPLY(x, m) (((x) * (m)) >> 23)
+#else     // (int)(x * a / 255. + .5)
+#define MULTIPLIER(a) ((a) * 65793UL)
+#define PREMULTIPLY(x, m) (((x) * (m) + (1UL << 23)) >> 24)
+#endif
+
+static void ApplyAlphaMultiply(uint8_t* rgba, int alpha_1rst,
+                               int w, int h, int stride) {
+  while (h-- > 0) {
+    uint8_t* const rgb = rgba + (alpha_1rst ? 1 : 0);
+    const uint8_t* const alpha = rgba + (alpha_1rst ? 0 : 3);
+    int i;
+    for (i = 0; i < w; ++i) {
+      const uint32_t a = alpha[4 * i];
+      if (a != 0xff) {
+        const uint32_t mult = MULTIPLIER(a);
+        rgb[4 * i + 0] = PREMULTIPLY(rgb[4 * i + 0], mult);
+        rgb[4 * i + 1] = PREMULTIPLY(rgb[4 * i + 1], mult);
+        rgb[4 * i + 2] = PREMULTIPLY(rgb[4 * i + 2], mult);
+      }
+    }
+    rgba += stride;
+  }
+}
+#undef MULTIPLIER
+#undef PREMULTIPLY
+
+// rgbA4444
+
+#define MULTIPLIER(a)  ((a) * 0x11)
+#define PREMULTIPLY(x, m) (((x) * (m)) >> 12)
+
+static WEBP_INLINE uint8_t dither_hi(uint8_t x) { return (x & 0xf0) | (x >> 4); }
+static WEBP_INLINE uint8_t dither_lo(uint8_t x) { return (x & 0x0f) | (x << 4); }
+
+static void ApplyAlphaMultiply4444(uint8_t* rgba4444,
+                                   int w, int h, int stride) {
+  while (h-- > 0) {
+    int i;
+    for (i = 0; i < w; ++i) {
+      const uint8_t a = dither_lo(rgba4444[2 * i + 1]);
+      const uint32_t mult = MULTIPLIER(a);
+      const uint8_t r = PREMULTIPLY(dither_hi(rgba4444[2 * i + 0]), mult);
+      const uint8_t g = PREMULTIPLY(dither_lo(rgba4444[2 * i + 0]), mult);
+      const uint8_t b = PREMULTIPLY(dither_hi(rgba4444[2 * i + 1]), mult);
+      rgba4444[2 * i + 0] = (r & 0xf0) | (g & 0x0f);
+      rgba4444[2 * i + 1] = (b & 0xf0) | a;
+    }
+    rgba4444 += stride;
+  }
+}
+#undef MULTIPLIER
+#undef PREMULTIPLY
+
+void (*WebPApplyAlphaMultiply)(uint8_t*, int, int, int, int)
+    = ApplyAlphaMultiply;
+void (*WebPApplyAlphaMultiply4444)(uint8_t*, int, int, int)
+    = ApplyAlphaMultiply4444;
+
 //------------------------------------------------------------------------------
 // Main call
 
@@ -216,16 +272,8 @@ void WebPInitUpsamplers(void) {
   WebPUpsamplers[MODE_RGBA_4444] = UpsampleRgba4444LinePair;
   WebPUpsamplers[MODE_RGB_565]   = UpsampleRgb565LinePair;
 
-  WebPUpsamplersKeepAlpha[MODE_RGB]       = UpsampleRgbLinePair;
-  WebPUpsamplersKeepAlpha[MODE_RGBA]      = UpsampleRgbaKeepAlphaLinePair;
-  WebPUpsamplersKeepAlpha[MODE_BGR]       = UpsampleBgrLinePair;
-  WebPUpsamplersKeepAlpha[MODE_BGRA]      = UpsampleBgraKeepAlphaLinePair;
-  WebPUpsamplersKeepAlpha[MODE_ARGB]      = UpsampleArgbKeepAlphaLinePair;
-  WebPUpsamplersKeepAlpha[MODE_RGBA_4444] = UpsampleRgba4444KeepAlphaLinePair;
-  WebPUpsamplersKeepAlpha[MODE_RGB_565]   = UpsampleRgb565LinePair;
-
   // If defined, use CPUInfo() to overwrite some pointers with faster versions.
-  if (VP8GetCPUInfo) {
+  if (VP8GetCPUInfo != NULL) {
 #if defined(WEBP_USE_SSE2)
     if (VP8GetCPUInfo(kSSE2)) {
       WebPInitUpsamplersSSE2();
@@ -235,6 +283,26 @@ void WebPInitUpsamplers(void) {
 #endif  // FANCY_UPSAMPLING
 }
 
+void WebPInitPremultiply(void) {
+  WebPApplyAlphaMultiply = ApplyAlphaMultiply;
+  WebPApplyAlphaMultiply4444 = ApplyAlphaMultiply4444;
+
+#ifdef FANCY_UPSAMPLING
+  WebPUpsamplers[MODE_rgbA]      = UpsampleRgbaLinePair;
+  WebPUpsamplers[MODE_bgrA]      = UpsampleBgraLinePair;
+  WebPUpsamplers[MODE_Argb]      = UpsampleArgbLinePair;
+  WebPUpsamplers[MODE_rgbA_4444] = UpsampleRgba4444LinePair;
+
+  if (VP8GetCPUInfo != NULL) {
+#if defined(WEBP_USE_SSE2)
+    if (VP8GetCPUInfo(kSSE2)) {
+      WebPInitPremultiplySSE2();
+    }
+#endif
+  }
+#endif  // FANCY_UPSAMPLING
+}
+
 #if defined(__cplusplus) || defined(c_plusplus)
 }    // extern "C"
 #endif
diff --git a/src/dsp/upsampling_sse2.c b/src/dsp/upsampling_sse2.c
index 6116b0a5..93fc2cca 100644
--- a/src/dsp/upsampling_sse2.c
+++ b/src/dsp/upsampling_sse2.c
@@ -176,9 +176,6 @@ SSE2_UPSAMPLE_FUNC(UpsampleRgbLinePairSSE2,  VP8YuvToRgb,  3)
 SSE2_UPSAMPLE_FUNC(UpsampleBgrLinePairSSE2,  VP8YuvToBgr,  3)
 SSE2_UPSAMPLE_FUNC(UpsampleRgbaLinePairSSE2, VP8YuvToRgba, 4)
 SSE2_UPSAMPLE_FUNC(UpsampleBgraLinePairSSE2, VP8YuvToBgra, 4)
-// These two don't erase the alpha value
-SSE2_UPSAMPLE_FUNC(UpsampleRgbKeepAlphaLinePairSSE2, VP8YuvToRgb, 4)
-SSE2_UPSAMPLE_FUNC(UpsampleBgrKeepAlphaLinePairSSE2, VP8YuvToBgr, 4)
 
 #undef GET_M
 #undef PACK_AND_STORE
@@ -190,7 +187,6 @@ SSE2_UPSAMPLE_FUNC(UpsampleBgrKeepAlphaLinePairSSE2, VP8YuvToBgr, 4)
 //------------------------------------------------------------------------------
 
 extern WebPUpsampleLinePairFunc WebPUpsamplers[/* MODE_LAST */];
-extern WebPUpsampleLinePairFunc WebPUpsamplersKeepAlpha[/* MODE_LAST */];
 
 #endif  // FANCY_UPSAMPLING
 
@@ -200,11 +196,13 @@ void WebPInitUpsamplersSSE2(void) {
   WebPUpsamplers[MODE_RGBA] = UpsampleRgbaLinePairSSE2;
   WebPUpsamplers[MODE_BGR]  = UpsampleBgrLinePairSSE2;
   WebPUpsamplers[MODE_BGRA] = UpsampleBgraLinePairSSE2;
+#endif  // FANCY_UPSAMPLING
+}
 
-  WebPUpsamplersKeepAlpha[MODE_RGB]  = UpsampleRgbLinePairSSE2;
-  WebPUpsamplersKeepAlpha[MODE_RGBA] = UpsampleRgbKeepAlphaLinePairSSE2;
-  WebPUpsamplersKeepAlpha[MODE_BGR]  = UpsampleBgrLinePairSSE2;
-  WebPUpsamplersKeepAlpha[MODE_BGRA] = UpsampleBgrKeepAlphaLinePairSSE2;
+void WebPInitPremultiplySSE2(void) {
+#ifdef FANCY_UPSAMPLING
+  WebPUpsamplers[MODE_rgbA] = UpsampleRgbaLinePairSSE2;
+  WebPUpsamplers[MODE_bgrA] = UpsampleBgraLinePairSSE2;
 #endif  // FANCY_UPSAMPLING
 }
 
diff --git a/src/dsp/yuv.h b/src/dsp/yuv.h
index 5000b7d3..8f52acb2 100644
--- a/src/dsp/yuv.h
+++ b/src/dsp/yuv.h
@@ -99,6 +99,11 @@ static WEBP_INLINE void VP8YuvToRgba(uint8_t y, uint8_t u, uint8_t v,
   rgba[3] = 0xff;
 }
 
+static WEBP_INLINE uint32_t VP8Clip4Bits(uint8_t c) {
+  const uint32_t v = (c + 8) >> 4;
+  return (v > 15) ? 15 : v;
+}
+
 // Must be called before everything, to initialize the tables.
 void VP8YUVInit(void);
 
diff --git a/src/webp/decode.h b/src/webp/decode.h
index e3f0d346..dc8140b7 100644
--- a/src/webp/decode.h
+++ b/src/webp/decode.h
@@ -113,12 +113,35 @@ typedef enum { MODE_RGB = 0, MODE_RGBA = 1,
                MODE_BGR = 2, MODE_BGRA = 3,
                MODE_ARGB = 4, MODE_RGBA_4444 = 5,
                MODE_RGB_565 = 6,
+               // RGB-premultiplied transparent modes (alpha value is preserved)
+               MODE_rgbA = 7,
+               MODE_bgrA = 8,
+               MODE_Argb = 9,
+               MODE_rgbA_4444 = 10,
                // YUV modes must come after RGB ones.
-               MODE_YUV = 7, MODE_YUVA = 8,  // yuv 4:2:0
-               MODE_LAST = 9
+               MODE_YUV = 11, MODE_YUVA = 12,  // yuv 4:2:0
+               MODE_LAST = 13
              } WEBP_CSP_MODE;
 
-// Generic structure for describing the sample buffer.
+// Some useful macros:
+static WEBP_INLINE int WebPIsPremultipliedMode(WEBP_CSP_MODE mode) {
+  return (mode == MODE_rgbA || mode == MODE_bgrA || mode == MODE_Argb ||
+          mode == MODE_rgbA_4444);
+}
+
+static WEBP_INLINE int WebPIsAlphaMode(WEBP_CSP_MODE mode) {
+  return (mode == MODE_RGBA || mode == MODE_BGRA || mode == MODE_ARGB ||
+          mode == MODE_RGBA_4444 || mode == MODE_YUVA ||
+          WebPIsPremultipliedMode(mode));
+}
+
+static WEBP_INLINE int WebPIsRGBMode(WEBP_CSP_MODE mode) {
+  return (mode < MODE_YUV);
+}
+
+//------------------------------------------------------------------------------
+// WebPDecBuffer: Generic structure for describing the sample buffer.
+
 typedef struct {    // view as RGBA
   uint8_t* rgba;    // pointer to RGBA samples
   int stride;       // stride in bytes from one scanline to the next.
diff --git a/src/webp/decode_vp8.h b/src/webp/decode_vp8.h
index dde7fc94..7cbb2457 100644
--- a/src/webp/decode_vp8.h
+++ b/src/webp/decode_vp8.h
@@ -99,7 +99,9 @@ struct VP8Io {
   int use_scaling;
   int scaled_width, scaled_height;
 
-  // pointer to the alpha data (if present) corresponding to the rows
+  // If non NULL, pointer to the alpha data (if present) corresponding to the
+  // start of the current row (That is: it is pre-offset by mb_y and takes
+  // cropping into account).
   const uint8_t* a;
 };