diff --git a/jccoefct.c b/jccoefct.c
index 9c7162f9..0861d036 100644
--- a/jccoefct.c
+++ b/jccoefct.c
@@ -360,10 +360,9 @@ compress_trellis_pass (j_compress_ptr cinfo, JSAMPIMAGE input_buf)
   JBLOCKROW thisblockrow, lastblockrow;
   JBLOCKARRAY buffer_dst;
   
-  for (ci = 0; ci < cinfo->comps_in_scan; ci++) {    
+  for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
     c_derived_tbl actbl_data;
     c_derived_tbl *actbl = &actbl_data;
-    
     compptr = cinfo->cur_comp_info[ci];
 
     jpeg_make_c_derived_tbl(cinfo, FALSE, compptr->ac_tbl_no, &actbl);
@@ -398,7 +397,7 @@ compress_trellis_pass (j_compress_ptr cinfo, JSAMPIMAGE input_buf)
      */
     for (block_row = 0; block_row < block_rows; block_row++) {
       thisblockrow = buffer[block_row];
-      quantize_trellis(cinfo, actbl, thisblockrow, buffer_dst[block_row], blocks_across, cinfo->quant_tbl_ptrs[compptr->quant_tbl_no]);
+      quantize_trellis(cinfo, actbl, thisblockrow, buffer_dst[block_row], blocks_across, cinfo->quant_tbl_ptrs[compptr->quant_tbl_no], cinfo->norm_src[compptr->quant_tbl_no], cinfo->norm_coef[compptr->quant_tbl_no]);
       
       if (ndummy > 0) {
 	/* Create dummy blocks at the right edge of the image. */
@@ -435,6 +434,7 @@ compress_trellis_pass (j_compress_ptr cinfo, JSAMPIMAGE input_buf)
       }
     }
   }
+
   /* NB: compress_output will increment iMCU_row_num if successful.
    * A suspension return will result in redoing all the work above next time.
    */
diff --git a/jcdctmgr.c b/jcdctmgr.c
index 629b0053..833b55fd 100644
--- a/jcdctmgr.c
+++ b/jcdctmgr.c
@@ -576,7 +576,7 @@ static const float jpeg_lambda_weights_csf_luma[64] = {
 
 GLOBAL(void)
 quantize_trellis(j_compress_ptr cinfo, c_derived_tbl *actbl, JBLOCKROW coef_blocks, JBLOCKROW src, JDIMENSION num_blocks,
-                 JQUANT_TBL * qtbl)
+                 JQUANT_TBL * qtbl, double *norm_src, double *norm_coef)
 {
   int i, j, k;
   float accumulated_zero_dist[DCTSIZE2];
@@ -821,6 +821,15 @@ quantize_trellis(j_compress_ptr cinfo, c_derived_tbl *actbl, JBLOCKROW coef_bloc
     free(block_run_start);
     free(requires_eob);
   }
+  
+  if (cinfo->trellis_q_opt) {
+    for (bi = 0; bi < num_blocks; bi++) {
+      for (i = 1; i < DCTSIZE2; i++) {
+        norm_src[i] += src[bi][i] * coef_blocks[bi][i];
+        norm_coef[i] += 8 * coef_blocks[bi][i] * coef_blocks[bi][i];
+      }
+    }
+  }
 }
 
 /*
diff --git a/jchuff.h b/jchuff.h
index bdbc4a1e..ac469dab 100644
--- a/jchuff.h
+++ b/jchuff.h
@@ -50,4 +50,4 @@ EXTERN(void) jpeg_gen_optimal_table
 
 EXTERN(void) quantize_trellis
         JPP((j_compress_ptr cinfo, c_derived_tbl *actbl, JBLOCKROW coef_blocks, JBLOCKROW src, JDIMENSION num_blocks,
-                 JQUANT_TBL * qtbl));
+                 JQUANT_TBL * qtbl, double *norm_src, double *norm_coef));
diff --git a/jcmaster.c b/jcmaster.c
index f448ec41..bf3bd7ec 100644
--- a/jcmaster.c
+++ b/jcmaster.c
@@ -550,6 +550,16 @@ prepare_for_pass (j_compress_ptr cinfo)
     master->pub.call_pass_startup = FALSE;
     break;
   case trellis_pass:
+    if (master->pass_number%(cinfo->num_components*(cinfo->use_scans_in_trellis?4:2)) == 1 && cinfo->trellis_q_opt) {
+      int i, j;
+
+      for (i = 0; i < NUM_QUANT_TBLS; i++) {
+        for (j = 1; j < DCTSIZE2; j++) {
+          cinfo->norm_src[i][j] = 0.0;
+          cinfo->norm_coef[i][j] = 0.0;
+        }
+      }
+    }
     (*cinfo->entropy->start_pass) (cinfo, TRUE);
     (*cinfo->coef->start_pass) (cinfo, JBUF_REQUANT);
     master->pub.call_pass_startup = FALSE;
@@ -838,6 +848,20 @@ finish_pass_master (j_compress_ptr cinfo)
     break;
   case trellis_pass:
     master->pass_type = (cinfo->optimize_scans || master->pass_number < master->pass_number_scan_opt_base-1) ? huff_opt_pass : output_pass;
+      
+    if ((master->pass_number+1)%(cinfo->num_components*(cinfo->use_scans_in_trellis?4:2)) == 0 && cinfo->trellis_q_opt) {
+      int i, j;
+
+      for (i = 0; i < NUM_QUANT_TBLS; i++) {
+        for (j = 1; j < DCTSIZE2; j++) {
+          if (cinfo->norm_coef[i][j] != 0.0) {
+            int q = (int)(cinfo->norm_src[i][j] / cinfo->norm_coef[i][j] + 0.5);
+            if (q > 254) q = 254;
+            if (q < 1) q = 1;
+          }
+        }
+      }
+    }
     break;
   }
 
diff --git a/jcparam.c b/jcparam.c
index 883f9262..afe21026 100644
--- a/jcparam.c
+++ b/jcparam.c
@@ -433,6 +433,7 @@ jpeg_set_defaults (j_compress_ptr cinfo)
   cinfo->use_scans_in_trellis = FALSE;
   cinfo->trellis_freq_split = 8;
   cinfo->trellis_num_loops = 1;
+  cinfo->trellis_q_opt = FALSE;
 }
 
 
diff --git a/jpeglib.h b/jpeglib.h
index 23c3c846..5f6e2b0c 100644
--- a/jpeglib.h
+++ b/jpeglib.h
@@ -384,7 +384,11 @@ struct jpeg_compress_struct {
   boolean use_lambda_weight_tbl; /* TRUE=use lambda weighting table */
   boolean use_scans_in_trellis; /* TRUE=use scans in trellis optimization */
   boolean trellis_passes; /* TRUE=currently doing trellis-related passes */
+  boolean trellis_q_opt; /* TRUE=optimize quant table in trellis loop */
   
+  double norm_src[NUM_QUANT_TBLS][DCTSIZE2];
+  double norm_coef[NUM_QUANT_TBLS][DCTSIZE2];
+
   int trellis_freq_split; /* splitting point for frequency in trellis quantization */
   int trellis_num_loops; /* number of trellis loops */