diff --git a/src/bitstream.c b/src/bitstream.c
index 736f7677..6ae2ab79 100644
--- a/src/bitstream.c
+++ b/src/bitstream.c
@@ -88,8 +88,14 @@ void bitstream_init(bitstream* stream)
 void bitstream_alloc(bitstream* stream, uint32_t alloc)
 {
   stream->buffer = (uint8_t*)malloc(alloc);
+  stream->bufferlen = alloc;
   //Clear just to be sure
-  memset(stream->buffer,0,alloc);
+  bitstream_clear_buffer(stream);
+}
+
+void bitstream_clear_buffer(bitstream* stream)
+{
+  memset(stream->buffer,0,stream->bufferlen);
   stream->buffer_pos = 0;
 }
  
@@ -139,8 +145,6 @@ void bitstream_put(bitstream* stream, uint32_t data, uint8_t bits)
     stream->cur_bit = 0;
     if(stream->cur_byte==32)
     {
-        //We only have bytes 0..31
-        stream->cur_byte--;
         //Flush data out
         bitstream_flush(stream);
     }
@@ -151,10 +155,7 @@ void bitstream_put(bitstream* stream, uint32_t data, uint8_t bits)
   {
     stream->data[stream->cur_byte] |= (data<<(bitsleft-bits));
     stream->cur_bit+=bits;
-  }
- 
- 
- 
+  } 
 }
  
 /*
@@ -193,7 +194,7 @@ void bitstream_flush(bitstream* stream)
     if(stream->cur_byte)
     {
       /* Handle endianness issue */
-      for(i = 0; i < stream->cur_byte+1; i++)
+      for(i = 0; i < stream->cur_byte; i++)
       {
         /* "network" is big-endian */
         correct_endian = htonl(stream->data[i]);
diff --git a/src/bitstream.h b/src/bitstream.h
index b1def31b..867b2a7c 100644
--- a/src/bitstream.h
+++ b/src/bitstream.h
@@ -17,6 +17,7 @@ typedef struct
     FILE*    output;
     uint8_t* buffer;
     uint32_t buffer_pos;
+    uint32_t bufferlen;
 } bitstream;
 
 typedef struct
@@ -29,7 +30,8 @@ extern bitTable *g_exp_table;
 
 int floorLog2(unsigned int n);
  
-void bitstream_alloc(bitstream* stream, uint32_t alloc); 
+void bitstream_alloc(bitstream* stream, uint32_t alloc);
+void bitstream_clear_buffer(bitstream* stream);
 void bitstream_init(bitstream* stream); 
 void bitstream_put(bitstream* stream, uint32_t data, uint8_t bits); 
 /*
diff --git a/src/encoder.c b/src/encoder.c
index 125ca558..d17c5880 100644
--- a/src/encoder.c
+++ b/src/encoder.c
@@ -51,6 +51,8 @@ void encode_one_frame(encoder_control* encoder)
     bitstream_align(encoder->stream);
     bitstream_flush(encoder->stream);
     nal_write(encoder->output, encoder->stream->buffer, encoder->stream->buffer_pos, 1, NAL_SEQ_PARAMETER_SET, 0);
+        
+    bitstream_clear_buffer(encoder->stream);
 
     encode_pic_parameter_set(encoder);
     bitstream_align(encoder->stream);
@@ -72,8 +74,12 @@ void encode_pic_parameter_set(encoder_control* encoder)
   WRITE_U(encoder->stream, 0, 1, "sign_data_hiding_flag");
   WRITE_U(encoder->stream, 0, 1, "cabac_init_present_flag");
 
+  WRITE_U(encoder->stream, 0, 3, "num_ref_idx_l0_default_active_minus1");
+  WRITE_U(encoder->stream, 0, 3, "num_ref_idx_l1_default_active_minus1");
+  /*
   WRITE_UE(encoder->stream, 0, "num_ref_idx_l0_default_active_minus1");
   WRITE_UE(encoder->stream, 0, "num_ref_idx_l1_default_active_minus1");
+  */
   WRITE_SE(encoder->stream, 0, "pic_init_qp_minus26");
 
   WRITE_U(encoder->stream, 0, 1, "constrained_intra_pred_flag");
@@ -119,7 +125,7 @@ void encode_seq_parameter_set(encoder_control* encoder)
 
   WRITE_U(encoder->stream, 0, 1, "pcm_enabled_flag");
 
-  WRITE_U(encoder->stream, 0, 1, "qpprime_y_zero_transquant_bypass_flag");   
+  WRITE_U(encoder->stream, 0, 1, "qpprime_y_zero_transquant_bypass_flag");
 
   WRITE_UE(encoder->stream, 0, "log2_max_pic_order_cnt_lsb_minus4");
 
@@ -134,14 +140,14 @@ void encode_seq_parameter_set(encoder_control* encoder)
   WRITE_UE(encoder->stream, 0, "log2_min_transform_block_size_minus2");
   WRITE_UE(encoder->stream, 3, "log2_diff_max_min_transform_block_size");
 
+  WRITE_U(encoder->stream, 0, 1, "unknown_flag");
+
   WRITE_UE(encoder->stream, 2, "max_transform_hierarchy_depth_inter");
   WRITE_UE(encoder->stream, 2, "max_transform_hierarchy_depth_intra");
-
-
+  
   WRITE_U(encoder->stream, 0, 1, "scaling_list_enable_flag");
   WRITE_U(encoder->stream, 0, 1, "chroma_pred_from_luma_enabled_flag");
   WRITE_U(encoder->stream, 0, 1, "transform_skip_enabled_flag");
-	
 
   WRITE_U(encoder->stream, 0, 1, "deblocking_filter_in_aps_enabled_flag");
   WRITE_U(encoder->stream, 0, 1, "seq_loop_filter_across_slices_enabled_flag");
diff --git a/src/nal.c b/src/nal.c
index 5d25637c..fae25c87 100644
--- a/src/nal.c
+++ b/src/nal.c
@@ -27,10 +27,10 @@ void nal_write(FILE* output, uint8_t* buffer, uint32_t buffer_len, uint8_t nal_r
   uint8_t byte;
   uint32_t i;
   uint8_t zerocount=0;
-  uint8_t emulation_prevention_three_byte = 0x03;
-  uint8_t start_code_prefix_one_3bytes = 0x01;
-  uint8_t zero = 0x00;
-  
+  const uint8_t emulation_prevention_three_byte = 0x03;
+  const uint8_t start_code_prefix_one_3bytes = 0x01;
+  const uint8_t zero = 0x00;
+
   //start_code_prefix_one_3bytes
   fwrite(&zero, 1, 1, output);
   fwrite(&zero, 1, 1, output);
@@ -59,10 +59,12 @@ void nal_write(FILE* output, uint8_t* buffer, uint32_t buffer_len, uint8_t nal_r
     else
       zerocount = 0;
 
+    /* Write the actual data */
     fwrite(&buffer[i], 1, 1, output);
   }
+
   //If last byte was 0, add emulation_prevention_three_byte
-  //if(buffer[buffer_len-1] == 0)
-  //  fwrite(&emulation_prevention_three_byte, 1, 1, output);
+  if(buffer[buffer_len-1] == 0)
+    fwrite(&emulation_prevention_three_byte, 1, 1, output);
 
 }
\ No newline at end of file