Index: ffmpeg/libavcodec/huffyuv.c
===================================================================
--- ffmpeg/libavcodec/huffyuv.c	(revision 20545)
+++ ffmpeg/libavcodec/huffyuv.c	(working copy)
@@ -39,10 +39,12 @@
 #define B 3
 #define G 2
 #define R 1
+#define A 0
 #else
 #define B 0
 #define G 1
 #define R 2
+#define A 3
 #endif
 
 typedef enum Predictor{
@@ -61,7 +63,6 @@
     int bitstream_bpp;
     int version;
     int yuy2;                               //use yuy2 instead of 422P
-    int bgr32;                              //use bgr32 instead of bgr24
     int width, height;
     int flags;
     int context;
@@ -406,7 +407,8 @@
             s->temp[i]= av_malloc(s->width + 16);
         }
     }else{
-        s->temp[0]= av_malloc(4*s->width + 16);
+        int bytes = (s->bitstream_bpp==24) ? 3 : 4;
+        s->temp[0]= av_malloc(bytes*s->width + 16);
     }
 }
 
@@ -436,7 +438,6 @@
     avctx->coded_frame= &s->picture;
     s->interlaced= s->height > 288;
 
-s->bgr32=1;
 //if(avctx->extradata)
 //  printf("extradata:%X, extradata_size:%d\n", *(uint32_t*)avctx->extradata, avctx->extradata_size);
     if(avctx->extradata_size){
@@ -504,12 +505,10 @@
         }
         break;
     case 24:
+        avctx->pix_fmt = PIX_FMT_BGR24;
+        break;
     case 32:
-        if(s->bgr32){
-            avctx->pix_fmt = PIX_FMT_RGB32;
-        }else{
-            avctx->pix_fmt = PIX_FMT_BGR24;
-        }
+        avctx->pix_fmt = PIX_FMT_RGB32;
         break;
     default:
         assert(0);
@@ -820,7 +819,7 @@
 }
 #endif /* CONFIG_HUFFYUV_ENCODER || CONFIG_FFVHUFF_ENCODER */
 
-static av_always_inline void decode_bgr_1(HYuvContext *s, int count, int decorrelate, int alpha){
+static av_always_inline void decode_bgr_1_32(HYuvContext *s, int count, int decorrelate){
     int i;
     for(i=0; i<count; i++){
         int code = get_vlc2(&s->gb, s->vlc[3].table, VLC_BITS, 1);
@@ -835,22 +834,40 @@
             s->temp[0][4*i+G] = get_vlc2(&s->gb, s->vlc[1].table, VLC_BITS, 3);
             s->temp[0][4*i+R] = get_vlc2(&s->gb, s->vlc[2].table, VLC_BITS, 3);
         }
-        if(alpha)
-            get_vlc2(&s->gb, s->vlc[2].table, VLC_BITS, 3); //?!
+        s->temp[0][4*i+A] = get_vlc2(&s->gb, s->vlc[2].table, VLC_BITS, 3);
     }
 }
 
+static av_always_inline void decode_bgr_1_24(HYuvContext *s, int count, int decorrelate){
+    int i;
+    for(i=0; i<count; i++){
+        int code = get_vlc2(&s->gb, s->vlc[3].table, VLC_BITS, 1);
+        if(code != -1){
+            *(uint32_t*)&s->temp[0][3*i] = AV_RL32(&s->pix_bgr_map[code]);
+        }else if(decorrelate){
+            // B = 0, G = 1, R = 2 for BGR24 no matter the endianness
+            s->temp[0][3*i+1] = get_vlc2(&s->gb, s->vlc[1].table, VLC_BITS, 3);
+            s->temp[0][3*i+0] = get_vlc2(&s->gb, s->vlc[0].table, VLC_BITS, 3) + s->temp[0][3*i+1];
+            s->temp[0][3*i+2] = get_vlc2(&s->gb, s->vlc[2].table, VLC_BITS, 3) + s->temp[0][3*i+1];
+        }else{
+            s->temp[0][3*i+0] = get_vlc2(&s->gb, s->vlc[0].table, VLC_BITS, 3);
+            s->temp[0][3*i+1] = get_vlc2(&s->gb, s->vlc[1].table, VLC_BITS, 3);
+            s->temp[0][3*i+2] = get_vlc2(&s->gb, s->vlc[2].table, VLC_BITS, 3);
+        }
+    }
+}
+
 static void decode_bgr_bitstream(HYuvContext *s, int count){
     if(s->decorrelate){
         if(s->bitstream_bpp==24)
-            decode_bgr_1(s, count, 1, 0);
+            decode_bgr_1_24(s, count, 1);
         else
-            decode_bgr_1(s, count, 1, 1);
+            decode_bgr_1_32(s, count, 1);
     }else{
         if(s->bitstream_bpp==24)
-            decode_bgr_1(s, count, 0, 0);
+            decode_bgr_1_24(s, count, 0);
         else
-            decode_bgr_1(s, count, 0, 1);
+            decode_bgr_1_32(s, count, 0);
     }
 }
 
@@ -1116,32 +1133,38 @@
         }
     }else{
         int y;
-        int leftr, leftg, leftb;
+        int leftr, leftg, leftb, lefta;
         const int last_line= (height-1)*p->linesize[0];
 
         if(s->bitstream_bpp==32){
-            skip_bits(&s->gb, 8);
+            lefta= p->data[0][last_line+A]= get_bits(&s->gb, 8);
             leftr= p->data[0][last_line+R]= get_bits(&s->gb, 8);
             leftg= p->data[0][last_line+G]= get_bits(&s->gb, 8);
             leftb= p->data[0][last_line+B]= get_bits(&s->gb, 8);
         }else{
-            leftr= p->data[0][last_line+R]= get_bits(&s->gb, 8);
-            leftg= p->data[0][last_line+G]= get_bits(&s->gb, 8);
-            leftb= p->data[0][last_line+B]= get_bits(&s->gb, 8);
+            leftr= p->data[0][last_line+2]= get_bits(&s->gb, 8);
+            leftg= p->data[0][last_line+1]= get_bits(&s->gb, 8);
+            leftb= p->data[0][last_line+0]= get_bits(&s->gb, 8);
             skip_bits(&s->gb, 8);
         }
 
-        if(s->bgr32){
             switch(s->predictor){
             case LEFT:
             case PLANE:
                 decode_bgr_bitstream(s, width-1);
-                s->dsp.add_hfyu_left_prediction_bgr32(p->data[0] + last_line+4, s->temp[0], width-1, &leftr, &leftg, &leftb);
+                if(s->bitstream_bpp==32)
+                    s->dsp.add_hfyu_left_prediction_bgr32(p->data[0] + last_line+4, s->temp[0], width-1, &leftr, &leftg, &leftb, &lefta);
+                else
+                    s->dsp.add_hfyu_left_prediction_bgr24(p->data[0] + last_line+3, s->temp[0], width-1, &leftr, &leftg, &leftb);
 
+
                 for(y=s->height-2; y>=0; y--){ //Yes it is stored upside down.
                     decode_bgr_bitstream(s, width);
 
-                    s->dsp.add_hfyu_left_prediction_bgr32(p->data[0] + p->linesize[0]*y, s->temp[0], width, &leftr, &leftg, &leftb);
+                    if(s->bitstream_bpp==32)
+                        s->dsp.add_hfyu_left_prediction_bgr32(p->data[0] + p->linesize[0]*y, s->temp[0], width, &leftr, &leftg, &leftb, &lefta);
+                    else
+                        s->dsp.add_hfyu_left_prediction_bgr24(p->data[0] + p->linesize[0]*y, s->temp[0], width, &leftr, &leftg, &leftb);
                     if(s->predictor == PLANE){
                         if((y&s->interlaced)==0 && y<s->height-1-s->interlaced){
                             s->dsp.add_bytes(p->data[0] + p->linesize[0]*y,
@@ -1154,11 +1177,6 @@
             default:
                 av_log(avctx, AV_LOG_ERROR, "prediction type not supported!\n");
             }
-        }else{
-
-            av_log(avctx, AV_LOG_ERROR, "BGR24 output is not implemented yet\n");
-            return -1;
-        }
     }
     emms_c();
 
Index: ffmpeg/libavcodec/dsputil.c
===================================================================
--- ffmpeg/libavcodec/dsputil.c	(revision 20545)
+++ ffmpeg/libavcodec/dsputil.c	(working copy)
@@ -3627,39 +3627,68 @@
     return acc;
 }
 
+static inline void add_hfyu_left_prediction_bgr24_c(uint8_t *dst, const uint8_t *src, int w, int *red, int *green, int *blue){
+    int i;
+    int r,g,b;
+    r= *red;
+    g= *green;
+    b= *blue;
+
+    for(i=0; i<w; i++){
+        b+= src[3*i+0];
+        g+= src[3*i+1];
+        r+= src[3*i+2];
+
+        dst[3*i+0]= b;
+        dst[3*i+1]= g;
+        dst[3*i+2]= r;
+    }
+
+    *red= r;
+    *green= g;
+    *blue= b;
+}
+
 #if HAVE_BIGENDIAN
 #define B 3
 #define G 2
 #define R 1
+#define A 0
 #else
 #define B 0
 #define G 1
 #define R 2
+#define A 3
 #endif
-static void add_hfyu_left_prediction_bgr32_c(uint8_t *dst, const uint8_t *src, int w, int *red, int *green, int *blue){
+static void add_hfyu_left_prediction_bgr32_c(uint8_t *dst, const uint8_t *src, int w, int *red, int *green, int *blue, int *alpha){
     int i;
-    int r,g,b;
+    int r,g,b,a;
     r= *red;
     g= *green;
     b= *blue;
+    a= *alpha;
 
     for(i=0; i<w; i++){
         b+= src[4*i+B];
         g+= src[4*i+G];
         r+= src[4*i+R];
+        a+= src[4*i+A];
 
         dst[4*i+B]= b;
         dst[4*i+G]= g;
         dst[4*i+R]= r;
+        dst[4*i+A]= a;
     }
 
     *red= r;
     *green= g;
     *blue= b;
+    *alpha= a;
 }
 #undef B
 #undef G
 #undef R
+#undef A
 
 #define BUTTERFLY2(o1,o2,i1,i2) \
 o1= (i1)+(i2);\
@@ -4791,6 +4820,7 @@
     c->add_hfyu_median_prediction= add_hfyu_median_prediction_c;
     c->sub_hfyu_median_prediction= sub_hfyu_median_prediction_c;
     c->add_hfyu_left_prediction  = add_hfyu_left_prediction_c;
+    c->add_hfyu_left_prediction_bgr24 = add_hfyu_left_prediction_bgr24_c;
     c->add_hfyu_left_prediction_bgr32 = add_hfyu_left_prediction_bgr32_c;
     c->bswap_buf= bswap_buf;
 #if CONFIG_PNG_DECODER
Index: ffmpeg/libavcodec/dsputil.h
===================================================================
--- ffmpeg/libavcodec/dsputil.h	(revision 20545)
+++ ffmpeg/libavcodec/dsputil.h	(working copy)
@@ -350,7 +350,8 @@
     void (*sub_hfyu_median_prediction)(uint8_t *dst, const uint8_t *src1, const uint8_t *src2, int w, int *left, int *left_top);
     void (*add_hfyu_median_prediction)(uint8_t *dst, const uint8_t *top, const uint8_t *diff, int w, int *left, int *left_top);
     int  (*add_hfyu_left_prediction)(uint8_t *dst, const uint8_t *src, int w, int left);
-    void (*add_hfyu_left_prediction_bgr32)(uint8_t *dst, const uint8_t *src, int w, int *red, int *green, int *blue);
+    void (*add_hfyu_left_prediction_bgr24)(uint8_t *dst, const uint8_t *src, int w, int *red, int *green, int *blue);
+    void (*add_hfyu_left_prediction_bgr32)(uint8_t *dst, const uint8_t *src, int w, int *red, int *green, int *blue, int *alpha);
     /* this might write to dst[w] */
     void (*add_png_paeth_prediction)(uint8_t *dst, uint8_t *src, uint8_t *top, int w, int bpp);
     void (*bswap_buf)(uint32_t *dst, const uint32_t *src, int w);