diff --git a/src/dsp/upsampling.c b/src/dsp/upsampling.c
index 151a32ee..651274fc 100644
--- a/src/dsp/upsampling.c
+++ b/src/dsp/upsampling.c
@@ -153,25 +153,28 @@ WebPUpsampleLinePairFunc WebPGetLinePairConverter(int alpha_is_last) {
 // YUV444 converter
 
 #define YUV444_FUNC(FUNC_NAME, FUNC, XSTEP)                                    \
-static void FUNC_NAME(const uint8_t* y, const uint8_t* u, const uint8_t* v,    \
-                      uint8_t* dst, int len) {                                 \
+extern void FUNC_NAME(const uint8_t* y, const uint8_t* u, const uint8_t* v,    \
+                      uint8_t* dst, int len);                                  \
+void FUNC_NAME(const uint8_t* y, const uint8_t* u, const uint8_t* v,           \
+               uint8_t* dst, int len) {                                        \
   int i;                                                                       \
   for (i = 0; i < len; ++i) FUNC(y[i], u[i], v[i], &dst[i * XSTEP]);           \
 }
 
-YUV444_FUNC(Yuv444ToRgb,      VP8YuvToRgb,  3)
-YUV444_FUNC(Yuv444ToBgr,      VP8YuvToBgr,  3)
-YUV444_FUNC(Yuv444ToRgba,     VP8YuvToRgba, 4)
-YUV444_FUNC(Yuv444ToBgra,     VP8YuvToBgra, 4)
-YUV444_FUNC(Yuv444ToArgb,     VP8YuvToArgb, 4)
-YUV444_FUNC(Yuv444ToRgba4444, VP8YuvToRgba4444, 2)
-YUV444_FUNC(Yuv444ToRgb565,   VP8YuvToRgb565, 2)
+YUV444_FUNC(WebPYuv444ToRgbC,      VP8YuvToRgb,  3)
+YUV444_FUNC(WebPYuv444ToBgrC,      VP8YuvToBgr,  3)
+YUV444_FUNC(WebPYuv444ToRgbaC,     VP8YuvToRgba, 4)
+YUV444_FUNC(WebPYuv444ToBgraC,     VP8YuvToBgra, 4)
+YUV444_FUNC(WebPYuv444ToArgbC,     VP8YuvToArgb, 4)
+YUV444_FUNC(WebPYuv444ToRgba4444C, VP8YuvToRgba4444, 2)
+YUV444_FUNC(WebPYuv444ToRgb565C,   VP8YuvToRgb565, 2)
 
 #undef YUV444_FUNC
 
 WebPYUV444Converter WebPYUV444Converters[MODE_LAST];
 
 extern void WebPInitYUV444ConvertersMIPSdspR2(void);
+extern void WebPInitYUV444ConvertersSSE2(void);
 
 static volatile VP8CPUInfo upsampling_last_cpuinfo_used1 =
     (VP8CPUInfo)&upsampling_last_cpuinfo_used1;
@@ -179,19 +182,24 @@ static volatile VP8CPUInfo upsampling_last_cpuinfo_used1 =
 WEBP_TSAN_IGNORE_FUNCTION void WebPInitYUV444Converters(void) {
   if (upsampling_last_cpuinfo_used1 == VP8GetCPUInfo) return;
 
-  WebPYUV444Converters[MODE_RGB]       = Yuv444ToRgb;
-  WebPYUV444Converters[MODE_RGBA]      = Yuv444ToRgba;
-  WebPYUV444Converters[MODE_BGR]       = Yuv444ToBgr;
-  WebPYUV444Converters[MODE_BGRA]      = Yuv444ToBgra;
-  WebPYUV444Converters[MODE_ARGB]      = Yuv444ToArgb;
-  WebPYUV444Converters[MODE_RGBA_4444] = Yuv444ToRgba4444;
-  WebPYUV444Converters[MODE_RGB_565]   = Yuv444ToRgb565;
-  WebPYUV444Converters[MODE_rgbA]      = Yuv444ToRgba;
-  WebPYUV444Converters[MODE_bgrA]      = Yuv444ToBgra;
-  WebPYUV444Converters[MODE_Argb]      = Yuv444ToArgb;
-  WebPYUV444Converters[MODE_rgbA_4444] = Yuv444ToRgba4444;
+  WebPYUV444Converters[MODE_RGB]       = WebPYuv444ToRgbC;
+  WebPYUV444Converters[MODE_RGBA]      = WebPYuv444ToRgbaC;
+  WebPYUV444Converters[MODE_BGR]       = WebPYuv444ToBgrC;
+  WebPYUV444Converters[MODE_BGRA]      = WebPYuv444ToBgraC;
+  WebPYUV444Converters[MODE_ARGB]      = WebPYuv444ToArgbC;
+  WebPYUV444Converters[MODE_RGBA_4444] = WebPYuv444ToRgba4444C;
+  WebPYUV444Converters[MODE_RGB_565]   = WebPYuv444ToRgb565C;
+  WebPYUV444Converters[MODE_rgbA]      = WebPYuv444ToRgbaC;
+  WebPYUV444Converters[MODE_bgrA]      = WebPYuv444ToBgraC;
+  WebPYUV444Converters[MODE_Argb]      = WebPYuv444ToArgbC;
+  WebPYUV444Converters[MODE_rgbA_4444] = WebPYuv444ToRgba4444C;
 
   if (VP8GetCPUInfo != NULL) {
+#if defined(WEBP_USE_SSE2)
+    if (VP8GetCPUInfo(kSSE2)) {
+      WebPInitYUV444ConvertersSSE2();
+    }
+#endif
 #if defined(WEBP_USE_MIPS_DSP_R2)
     if (VP8GetCPUInfo(kMIPSdspR2)) {
       WebPInitYUV444ConvertersMIPSdspR2();
diff --git a/src/dsp/upsampling_sse2.c b/src/dsp/upsampling_sse2.c
index ce5c1725..b85808e2 100644
--- a/src/dsp/upsampling_sse2.c
+++ b/src/dsp/upsampling_sse2.c
@@ -201,6 +201,41 @@ WEBP_TSAN_IGNORE_FUNCTION void WebPInitUpsamplersSSE2(void) {
 
 #endif  // FANCY_UPSAMPLING
 
+//------------------------------------------------------------------------------
+
+extern WebPYUV444Converter WebPYUV444Converters[/* MODE_LAST */];
+extern void WebPInitYUV444ConvertersSSE2(void);
+
+#define YUV444_FUNC(FUNC_NAME, CALL, XSTEP) \
+extern void WebP##FUNC_NAME##C(const uint8_t* y, const uint8_t* u,             \
+                               const uint8_t* v, uint8_t* dst, int len);       \
+static void FUNC_NAME(const uint8_t* y, const uint8_t* u, const uint8_t* v,    \
+                      uint8_t* dst, int len) {                                 \
+  int i;                                                                       \
+  const int max_len = len & ~31;                                               \
+  for (i = 0; i < max_len; i += 32) CALL(y + i, u + i, v + i, dst + i * XSTEP);\
+  if (i < len) {  /* C-fallback */                                             \
+    WebP##FUNC_NAME##C(y + i, u + i, v + i, dst + i * XSTEP, len - i);         \
+  }                                                                            \
+}
+
+YUV444_FUNC(Yuv444ToRgba, VP8YuvToRgba32, 4);
+YUV444_FUNC(Yuv444ToBgra, VP8YuvToBgra32, 4);
+YUV444_FUNC(Yuv444ToRgb, VP8YuvToRgb32, 3);
+YUV444_FUNC(Yuv444ToBgr, VP8YuvToBgr32, 3);
+
+WEBP_TSAN_IGNORE_FUNCTION void WebPInitYUV444ConvertersSSE2(void) {
+  VP8YUVInitSSE2();
+  WebPYUV444Converters[MODE_RGBA] = Yuv444ToRgba;
+  WebPYUV444Converters[MODE_BGRA] = Yuv444ToBgra;
+  WebPYUV444Converters[MODE_RGB]  = Yuv444ToRgb;
+  WebPYUV444Converters[MODE_BGR]  = Yuv444ToBgr;
+}
+
+#else
+
+WEBP_DSP_INIT_STUB(WebPInitYUV444ConvertersSSE2)
+
 #endif  // WEBP_USE_SSE2
 
 #if !(defined(FANCY_UPSAMPLING) && defined(WEBP_USE_SSE2))
diff --git a/src/dsp/yuv.h b/src/dsp/yuv.h
index 8a47edd8..af435a5b 100644
--- a/src/dsp/yuv.h
+++ b/src/dsp/yuv.h
@@ -249,7 +249,6 @@ void VP8YUVInit(void);
 // to the binary size. Otherwise, they are initialized at run-time (small cost).
 #define WEBP_YUV_USE_SSE2_TABLES
 
-#if defined(FANCY_UPSAMPLING)
 // Process 32 pixels and store the result (24b or 32b per pixel) in *dst.
 void VP8YuvToRgba32(const uint8_t* y, const uint8_t* u, const uint8_t* v,
                     uint8_t* dst);
@@ -259,7 +258,6 @@ void VP8YuvToBgra32(const uint8_t* y, const uint8_t* u, const uint8_t* v,
                     uint8_t* dst);
 void VP8YuvToBgr32(const uint8_t* y, const uint8_t* u, const uint8_t* v,
                    uint8_t* dst);
-#endif  // FANCY_UPSAMPLING
 
 // Must be called to initialize tables before using the functions.
 void VP8YUVInitSSE2(void);
diff --git a/src/dsp/yuv_sse2.c b/src/dsp/yuv_sse2.c
index 47d99e73..2f7a1835 100644
--- a/src/dsp/yuv_sse2.c
+++ b/src/dsp/yuv_sse2.c
@@ -125,8 +125,6 @@ static WEBP_INLINE void YuvToBgrSSE2(uint8_t y, uint8_t u, uint8_t v,
 //-----------------------------------------------------------------------------
 // Convert spans of 32 pixels to various RGB formats for the fancy upsampler.
 
-#ifdef FANCY_UPSAMPLING
-
 void VP8YuvToRgba32(const uint8_t* y, const uint8_t* u, const uint8_t* v,
                     uint8_t* dst) {
   int n;
@@ -186,8 +184,6 @@ void VP8YuvToBgr32(const uint8_t* y, const uint8_t* u, const uint8_t* v,
   memcpy(dst + n * 3, tmp, 2 * 3);
 }
 
-#endif  // FANCY_UPSAMPLING
-
 //-----------------------------------------------------------------------------
 // Arbitrary-length row conversion functions