diff --git a/src/xpra/vpx/codec.pyx b/src/xpra/vpx/codec.pyx
index ff87fca7fb..da360519c1 100644
--- a/src/xpra/vpx/codec.pyx
+++ b/src/xpra/vpx/codec.pyx
@@ -16,12 +16,13 @@ ctypedef void vpx_image_t
 cdef extern from "vpxlib.h":
     vpx_codec_ctx_t* init_encoder(int width, int height)
     void clean_encoder(vpx_codec_ctx_t *context)
-    vpx_image_t* csc_image(vpx_codec_ctx_t *ctx, uint8_t *input, int stride)
+    vpx_image_t* csc_image_rgb2yuv(vpx_codec_ctx_t *ctx, uint8_t *input, int stride)
+    int csc_image_yuv2rgb(vpx_codec_ctx_t *ctx, uint8_t *input[3], int stride[3], uint8_t **out, int *outsz, int *outstride)
     int compress_image(vpx_codec_ctx_t *ctx, vpx_image_t *image, uint8_t **out, int *outsz) nogil
 
     vpx_codec_ctx_t* init_decoder(int width, int height)
     void clean_decoder(vpx_codec_ctx_t *context)
-    int decompress_image(vpx_codec_ctx_t *context, uint8_t *input, int size, uint8_t **out, int *outsize, int *outstride)
+    int decompress_image(vpx_codec_ctx_t *context, uint8_t *input, int size, uint8_t *(*out)[3], int *outsize, int (*outstride)[3])
 
 
 ENCODERS = {}
@@ -64,19 +65,42 @@ cdef class Decoder(xcoder):
             clean_decoder(self.context)
             self.context = NULL
 
+    def decompress_image_to_yuv(self, input):
+        cdef uint8_t *dout[3]
+        cdef int outsize
+        cdef int outstrides[3]
+        cdef unsigned char * buf = <uint8_t *> 0
+        cdef Py_ssize_t buf_len = 0
+        assert self.context!=NULL
+        PyObject_AsReadBuffer(input, <void **>&buf, &buf_len)
+        i = decompress_image(self.context, buf, buf_len, &dout, &outsize, &outstrides)
+        if i!=0:
+            return i, [0, 0, 0], ["", "", ""]
+        doutvY = (<char *>dout[0])[:self.height * outstrides[0]]
+        doutvU = (<char *>dout[1])[:self.height * outstrides[1]]
+        doutvV = (<char *>dout[2])[:self.height * outstrides[2]]
+        out = [doutvY, doutvU, doutvV]
+        strides = [outstrides[0], outstrides[1], outstrides[2]]
+        return  i, strides, out
+    
     def decompress_image_to_rgb(self, input):
+        cdef uint8_t *yuvplanes[3]
         cdef uint8_t *dout
         cdef int outsize
+        cdef int yuvstrides[3]
         cdef int outstride
         cdef unsigned char * buf = <uint8_t *> 0
         cdef Py_ssize_t buf_len = 0
         assert self.context!=NULL
         assert self.last_image==NULL
         PyObject_AsReadBuffer(input, <void **>&buf, &buf_len)
-        i = decompress_image(self.context, buf, buf_len, &dout, &outsize, &outstride)
-        self.last_image = dout
+        i = decompress_image(self.context, buf, buf_len, &yuvplanes, &outsize, &yuvstrides)
         if i!=0:
             return i, 0, ""
+        i = csc_image_yuv2rgb(self.context, yuvplanes, yuvstrides, &dout, &outsize, &outstride)
+        if i!=0:
+            return i, 0, ""
+        self.last_image = dout
         doutv = (<char *>dout)[:outsize]
         return  i, outstride, doutv
 
@@ -105,7 +129,7 @@ cdef class Encoder(xcoder):
         assert self.context!=NULL
         #colourspace conversion with gil held:
         PyObject_AsReadBuffer(input, <void **>&buf, &buf_len)
-        pic_in = csc_image(self.context, buf, rowstride)
+        pic_in = csc_image_rgb2yuv(self.context, buf, rowstride)
         assert pic_in!=NULL, "colourspace conversion failed"
         #actual compression (no gil):
         with nogil:
diff --git a/src/xpra/vpx/vpxlib.c b/src/xpra/vpx/vpxlib.c
index 10272c9af2..b5d5dc39dd 100644
--- a/src/xpra/vpx/vpxlib.c
+++ b/src/xpra/vpx/vpxlib.c
@@ -99,7 +99,7 @@ void clean_decoder(struct vpx_context *ctx)
 	free(ctx);
 }
 
-vpx_image_t* csc_image(struct vpx_context *ctx, const uint8_t *in, int stride)
+vpx_image_t* csc_image_rgb2yuv(struct vpx_context *ctx, const uint8_t *in, int stride)
 {
 	vpx_image_t *image = malloc(sizeof(vpx_image_t));
 	if (!vpx_img_alloc(image, VPX_IMG_FMT_I420, ctx->width, ctx->height, 1)) {
@@ -144,7 +144,26 @@ int compress_image(struct vpx_context *ctx, vpx_image_t *image, uint8_t **out, i
 	return 0;
 }
 
-int decompress_image(struct vpx_context *ctx, uint8_t *in, int size, uint8_t **out, int *outsize, int *outstride)
+int csc_image_yuv2rgb(struct vpx_context *ctx, uint8_t *in[3], const int stride[3], uint8_t **out, int *outsz, int *outstride)
+{
+	int aligned_width;
+	uint8_t *dst[4] = { malloc(ctx->height * ctx->width * 3), NULL, NULL, NULL };
+	int dststride[4] = { ctx->width * 3, 0, 0, 0 };
+	
+	if (!ctx->yuv2rgb)
+		return 1;
+	
+	sws_scale(ctx->yuv2rgb, in, stride, 0, ctx->height, dst, dststride);
+	
+	/* Output (must be freed!) */
+	*out = dst[0];
+	*outsz = dststride[0] * ctx->height;
+	*outstride = dststride[0];
+
+	return 0;
+}
+
+int decompress_image(struct vpx_context *ctx, uint8_t *in, int size, uint8_t *(*out)[3], int *outsize, int (*outstride)[3])
 {
 	vpx_image_t      *img;
 	int frame_sz = size;
@@ -152,7 +171,6 @@ int decompress_image(struct vpx_context *ctx, uint8_t *in, int size, uint8_t **o
 	uint8_t* frame = in;
 	int outstrides[4];
 	uint8_t* outs[4];
-	int stride = 0;
 	int i = 0;
 
 	if (vpx_codec_decode(&ctx->codec, frame, frame_sz, NULL, 0)) {
@@ -164,19 +182,13 @@ int decompress_image(struct vpx_context *ctx, uint8_t *in, int size, uint8_t **o
 		codec_error(&ctx->codec, "vpx_codec_get_frame");
 		return -1;
 	}
-	for (i=0; i<4; i++)
-		stride += img->stride[i];
-	*outsize = stride * img->h;
-
-	*out = malloc(*outsize);
-	for (i=0; i<4; i++) {
-		outstrides[i] = img->w*3;
-		outs[i] = *out;
+	
+	*outsize = 0;
+	for (i = 0; i < 3; i++) {
+		(*out)[i] = img->planes[i];
+		(*outstride)[i] = img->stride[i];
+		*outsize += img->stride[i] * img->h;
 	}
-	sws_scale(ctx->yuv2rgb, img->planes, img->stride, 0, img->h, outs, outstrides);
-	stride = 0;
-	for (i=0; i<4; i++)
-		stride += img->stride[i];
-	*outstride = stride;
+
 	return 0;
 }
diff --git a/src/xpra/vpx/vpxlib.h b/src/xpra/vpx/vpxlib.h
index 2f5f659d53..735f1d8922 100644
--- a/src/xpra/vpx/vpxlib.h
+++ b/src/xpra/vpx/vpxlib.h
@@ -27,7 +27,17 @@ void clean_decoder(struct vpx_context *ctx);
  @param stride: Input stride (size is taken from context).
  @return: the converted picture.
 */
-vpx_image_t* csc_image(struct vpx_context *ctx, const uint8_t *in, int stride);
+vpx_image_t* csc_image_rgb2yuv(struct vpx_context *ctx, const uint8_t *in, int stride);
+
+/** Colorspace conversion.
+ @param in: Input picture (3 planes).
+ @param stride: Input strides (3 planes).
+ @param out: Will be set to point to the output data in packed RGB24 format. Must be freed after use by calling free().
+ @param outsz: Will be set to the size of the output buffer.
+ @param outstride: Output stride.
+ @return non zero on error.
+*/
+int csc_image_yuv2rgb(struct vpx_context *ctx, uint8_t *in[3], const int stride[3], uint8_t **out, int *outsz, int *outstride);
 
 /** Compress an image using the given context.
  @param pic_in: the input image, as returned by csc_image
@@ -40,7 +50,8 @@ int compress_image(struct vpx_context *ctx, vpx_image_t *image, uint8_t **out, i
 /** Decompress an image using the given context.
  @param in: Input buffer, format is H264.
  @param size: Input size.
- @param out: Will be set to point to the output data in RGB24 format.
- @param outstride: Output stride.
+ @param out: Will be filled to point to the output data in planar YUV420 format (3 planes). This data will be freed automatically upon next call to the decoder. 
+ @param outsize: Output size.
+ @param outstride: Output strides (3 planes).
 */
-int decompress_image(struct vpx_context *ctx, uint8_t *in, int size, uint8_t **out, int *outsize, int *outstride);
+int decompress_image(struct vpx_context *ctx, uint8_t *in, int size, uint8_t *(*out)[3], int *outsize, int (*outstride)[3]);
diff --git a/src/xpra/x264/x264lib.h b/src/xpra/x264/x264lib.h
index 9cb639a48d..33bc74a5c3 100644
--- a/src/xpra/x264/x264lib.h
+++ b/src/xpra/x264/x264lib.h
@@ -35,7 +35,6 @@ void clean_decoder(struct x264lib_ctx *);
 x264_picture_t *csc_image_rgb2yuv(struct x264lib_ctx *ctx, const uint8_t *in, int stride);
 
 /** Colorspace conversion.
- * Note: you must call compress_image to free the image buffer.
  @param in: Input picture (3 planes).
  @param stride: Input strides (3 planes).
  @param out: Will be set to point to the output data in packed RGB24 format. Must be freed after use by calling free().