diff --git a/src/cu.h b/src/cu.h
index 751a483c..a2e2234c 100644
--- a/src/cu.h
+++ b/src/cu.h
@@ -131,6 +131,9 @@ typedef struct
   uint8_t log2_width : 3;
   uint8_t log2_height : 3;
 
+  uint8_t log2_chroma_width : 3;
+  uint8_t log2_chroma_height : 3;
+
   uint16_t cbf;
 
   uint8_t root_cbf;
@@ -150,11 +153,14 @@ typedef struct
   uint8_t mts_last_scan_pos : 1;
 
   uint8_t violates_lfnst_constrained_luma : 1;
-  uint8_t violates_lfnst_constrained_chroma;
+  uint8_t violates_lfnst_constrained_chroma : 1;
   uint8_t lfnst_last_scan_pos : 1;
   uint8_t lfnst_idx : 2;
   uint8_t cr_lfnst_idx : 2;
 
+  uint8_t luma_deblocking : 2;
+  uint8_t chroma_deblocking : 2;
+
   union {
     struct {
       int8_t mode;
diff --git a/src/filter.c b/src/filter.c
index 2f0b6a1c..5605006b 100644
--- a/src/filter.c
+++ b/src/filter.c
@@ -269,6 +269,7 @@ static bool is_tu_boundary(
   int32_t x,
   int32_t y,
   edge_dir dir,
+  color_t color,
   enum uvg_tree_type tree_type)
 {
   x >>= tree_type == UVG_CHROMA_T;
@@ -276,13 +277,13 @@ static bool is_tu_boundary(
   // if (x & 3 || y & 3) return false;
   const cu_info_t *const scu =
     uvg_cu_array_at_const(tree_type != UVG_CHROMA_T ? state->tile->frame->cu_array : state->tile->frame->chroma_cu_array, x, y);
-  const int tu_width = MIN(TR_MAX_WIDTH, 1 << scu->log2_width);
-  const int tu_height = MIN(TR_MAX_WIDTH, 1 << scu->log2_height);
 
   if (dir == EDGE_HOR) {
-    return (y & (tu_height - 1)) == 0;
+    return color == COLOR_Y ? scu->luma_deblocking & EDGE_HOR :
+                              scu->chroma_deblocking & EDGE_HOR;
   } else {
-    return (x & (tu_width - 1)) == 0;
+    return color == COLOR_Y ? scu->luma_deblocking & EDGE_VER :
+                              scu->chroma_deblocking & EDGE_VER;
   }
 }
 
@@ -321,9 +322,9 @@ static bool is_pu_boundary(const encoder_state_t *const state,
 static bool is_on_8x8_grid(int x, int y, edge_dir dir)
 {
   if (dir == EDGE_HOR) {
-    return (y & 7) == 0 && (x & 2) == 0;
+    return (y & 7) == 0;
   } else {
-    return (x & 7) == 0 && (y & 2) == 0;
+    return (x & 7) == 0;
   }
 }
 
@@ -603,10 +604,10 @@ static INLINE void get_max_filter_length(uint8_t *filt_len_P, uint8_t *filt_len_
   bool transform_edge_4x4[2] = { false, false };
   bool transform_edge_8x8[2] = { false, false };
   
-  if (pos >= 4) transform_edge_4x4[0] = is_tu_boundary(state, x - x_mul * 4, y - y_mul * 4, dir, tree_type);
-  if (pos >= 8) transform_edge_8x8[0] = is_tu_boundary(state, x - x_mul * 8, y - y_mul * 8, dir, tree_type);
-  if (pos + 4 < len) transform_edge_4x4[1] = is_tu_boundary(state, x + x_mul * 4, y + y_mul * 4, dir, tree_type);
-  if (pos + 8 < len) transform_edge_8x8[1] = is_tu_boundary(state, x + x_mul * 8, y + y_mul * 8, dir, tree_type);
+  if (pos >= 4) transform_edge_4x4[0] = is_tu_boundary(state, x - x_mul * 4, y - y_mul * 4, dir, comp, tree_type);
+  if (pos >= 8) transform_edge_8x8[0] = is_tu_boundary(state, x - x_mul * 8, y - y_mul * 8, dir, comp, tree_type);
+  if (pos + 4 < len) transform_edge_4x4[1] = is_tu_boundary(state, x + x_mul * 4, y + y_mul * 4, dir, comp, tree_type);
+  if (pos + 8 < len) transform_edge_8x8[1] = is_tu_boundary(state, x + x_mul * 8, y + y_mul * 8, dir, comp, tree_type);
 
   if (comp == COLOR_Y) {
     if (tu_size_P_side <= 4 || tu_size_Q_side <= 4){
@@ -1066,18 +1067,18 @@ static void filter_deblock_edge_chroma(encoder_state_t * const state,
       uint8_t max_filter_length_P = 0;
       uint8_t max_filter_length_Q = 0;
       
-      const int cu_width = 1 << (cu_q->log2_width - (tree_type != UVG_CHROMA_T));
-      const int cu_height = 1 << (cu_q->log2_height - (tree_type != UVG_CHROMA_T));
+      const int cu_width = 1 << (cu_q->log2_chroma_width );
+      const int cu_height = 1 << (cu_q->log2_chroma_height);
       const int pu_size = dir == EDGE_HOR ? cu_height : cu_width;
       const int pu_pos = dir == EDGE_HOR ? y_coord : x_coord;
 
 
       const int tu_size_p_side = dir == EDGE_HOR ? 
-        MIN(1 << (cu_p->log2_height - (tree_type != UVG_CHROMA_T)), TR_MAX_WIDTH) :
-        MIN(1 << (cu_p->log2_width - (tree_type != UVG_CHROMA_T)), TR_MAX_WIDTH);
+        MIN(1 << (cu_p->log2_chroma_height), TR_MAX_WIDTH) :
+        MIN(1 << (cu_p->log2_chroma_width), TR_MAX_WIDTH);
       const int tu_size_q_side = dir == EDGE_HOR ?
-        MIN(1 << (cu_q->log2_height - (tree_type != UVG_CHROMA_T)), TR_MAX_WIDTH) :
-        MIN(1 << (cu_q->log2_width - (tree_type != UVG_CHROMA_T)), TR_MAX_WIDTH);
+        MIN(1 << (cu_q->log2_chroma_height ), TR_MAX_WIDTH) :
+        MIN(1 << (cu_q->log2_chroma_width ), TR_MAX_WIDTH);
 
       get_max_filter_length(&max_filter_length_P, &max_filter_length_Q, state, x_coord, y_coord,
                             dir, tu_boundary, tu_size_p_side, tu_size_q_side,
@@ -1216,11 +1217,12 @@ static void filter_deblock_unit(
   // Chroma pixel coordinates.
   const int32_t x_c = x >> 1;
   const int32_t y_c = y >> 1;
-  if (state->encoder_control->chroma_format != UVG_CSP_400 && 
-    (is_on_8x8_grid(x_c, y_c, dir && (x_c + 4) % 32)
-     || (x == state->tile->frame->width - 8 && dir == 1 && y_c % 8 == 0)) 
+  if (state->encoder_control->chroma_format != UVG_CSP_400 &&
+    is_tu_boundary(state, x, y, dir, COLOR_UV, tree_type)
+    && (is_on_8x8_grid(x_c, y_c, dir == EDGE_HOR && (x_c + 4) % 32 ? EDGE_HOR : EDGE_VER)
+     || (x == state->tile->frame->width - 8 && dir == EDGE_HOR && y_c % 8 == 0)) 
     && tree_type != UVG_LUMA_T) {
-    filter_deblock_edge_chroma(state, x_c, y_c, length, dir, tu_boundary, tree_type);
+    filter_deblock_edge_chroma(state, x_c, y_c, 2, dir, tu_boundary, tree_type);
   }
 }
 
@@ -1250,11 +1252,11 @@ static void filter_deblock_lcu_inside(encoder_state_t * const state,
 
   for (int edge_y = y; edge_y < end_y; edge_y += 4) {
     for (int edge_x = x; edge_x < end_x; edge_x += 4) {
-      bool tu_boundary = is_tu_boundary(state, edge_x, edge_y, dir, luma_tree);
+      bool tu_boundary = is_tu_boundary(state, edge_x, edge_y, dir, COLOR_Y, luma_tree);
       if (tu_boundary || is_pu_boundary(state, edge_x, edge_y, dir)) {
         filter_deblock_unit(state, edge_x, edge_y, 4, 4, dir, tu_boundary, edge_x < x, luma_tree);
       }
-      if(chroma_tree == UVG_CHROMA_T && is_tu_boundary(state, edge_x, edge_y, dir, chroma_tree)) {
+      if(chroma_tree == UVG_CHROMA_T && is_tu_boundary(state, edge_x, edge_y, dir, COLOR_UV, chroma_tree)) {
         filter_deblock_unit(state, edge_x, edge_y, 4, 4, dir, tu_boundary, edge_x < x, chroma_tree);        
       }
     }
@@ -1281,7 +1283,7 @@ static void filter_deblock_lcu_rightmost(encoder_state_t * const state,
   for (int x = x_px - 8; x < x_px; x += 4) {
     for (int y = y_px; y < end; y += 4) {
       // The top edge of the whole frame is not filtered.
-      bool tu_boundary = is_tu_boundary(state, x, y, EDGE_HOR, luma_tree);
+      bool tu_boundary = is_tu_boundary(state, x, y, EDGE_HOR, COLOR_Y, luma_tree);
       if (y > 0 && (tu_boundary || is_pu_boundary(state, x, y, EDGE_HOR))) {
         filter_deblock_edge_luma(state, x, y, 4, EDGE_HOR, tu_boundary);
       }
@@ -1292,13 +1294,15 @@ static void filter_deblock_lcu_rightmost(encoder_state_t * const state,
   if (state->encoder_control->chroma_format != UVG_CSP_400) {
     const int x_px_c = x_px >> 1;
     const int y_px_c = y_px >> 1;
-    const int x_c = x_px_c - 4;
-    const int end_c = MIN(y_px_c + LCU_WIDTH_C, state->tile->frame->height >> 1);
-    for (int y_c = y_px_c; y_c < end_c; y_c += 8) {
-      // The top edge of the whole frame is not filtered.
-      bool tu_boundary = is_tu_boundary(state, x_c << 1, y_c << 1, EDGE_HOR, chroma_tree);
-      if (y_c > 0 && (tu_boundary || is_pu_boundary(state, x_c << 1, y_c << 1, EDGE_HOR))) {
-        filter_deblock_edge_chroma(state, x_c , y_c, 4, EDGE_HOR, tu_boundary, chroma_tree);
+    int x_c = x_px_c - 4;
+    const int end_c_y = MIN(y_px_c + LCU_WIDTH_C, state->tile->frame->height >> 1);
+    for(; x_c < x_px_c; x_c += 2) {
+      for (int y_c = y_px_c; y_c < end_c_y; y_c += 8) {
+        // The top edge of the whole frame is not filtered.
+        bool tu_boundary = is_tu_boundary(state, x_c << 1, y_c << 1, EDGE_HOR, COLOR_UV, chroma_tree);
+        if (y_c > 0 && (tu_boundary || is_pu_boundary(state, x_c << 1, y_c << 1, EDGE_HOR))) {
+          filter_deblock_edge_chroma(state, x_c , y_c, 2, EDGE_HOR, tu_boundary, chroma_tree);
+        }
       }
     }
   }
diff --git a/src/filter.h b/src/filter.h
index 0d98eedd..2db9c871 100644
--- a/src/filter.h
+++ b/src/filter.h
@@ -46,8 +46,8 @@
  * \brief Edge direction.
  */
 typedef enum edge_dir {
-  EDGE_VER = 0, // vertical
-  EDGE_HOR = 1, // horizontal
+  EDGE_VER = 1, // vertical
+  EDGE_HOR = 2, // horizontal
 } edge_dir;
 
 
diff --git a/src/search.c b/src/search.c
index 74b3760f..18c721c6 100644
--- a/src/search.c
+++ b/src/search.c
@@ -39,6 +39,7 @@
 #include "cu.h"
 #include "encoder.h"
 #include "encode_coding_tree.h"
+#include "filter.h"
 #include "imagelist.h"
 #include "inter.h"
 #include "intra.h"
@@ -253,6 +254,10 @@ static void work_tree_copy_up(
         to_cu->intra.mode_chroma = from_cu->intra.mode_chroma;
         to_cu->joint_cb_cr = from_cu->joint_cb_cr;
         to_cu->cr_lfnst_idx = from_cu->cr_lfnst_idx;
+        to_cu->chroma_deblocking = from_cu->chroma_deblocking;
+        to_cu->log2_chroma_width = from_cu->log2_chroma_width;
+        to_cu->log2_chroma_height = from_cu->log2_chroma_height;
+
         cbf_copy(&to_cu->cbf, from_cu->cbf, COLOR_U);
         cbf_copy(&to_cu->cbf, from_cu->cbf, COLOR_V);
       }
@@ -282,6 +287,9 @@ static void lcu_fill_cu_info(lcu_t *lcu, int x_local, int y_local, int width, in
       to->log2_height = cu->log2_height;
       to->log2_width = cu->log2_width;
 
+      to->log2_chroma_height = cu->log2_chroma_height;
+      to->log2_chroma_width = cu->log2_chroma_width;
+
       if (cu->type == CU_INTRA) {
         to->intra.mode        = cu->intra.mode;
         to->intra.mode_chroma = cu->intra.mode_chroma;
@@ -315,14 +323,37 @@ static void lcu_fill_chroma_cu_info(lcu_t *lcu, const cu_loc_t * const cu_loc)
       cu->intra.mode_chroma = bottom_right->intra.mode_chroma;
       cu->joint_cb_cr       = bottom_right->joint_cb_cr;
       cu->cr_lfnst_idx      = bottom_right->cr_lfnst_idx;
+      cu->log2_chroma_height = bottom_right->log2_chroma_height;
+      cu->log2_chroma_width = bottom_right->log2_chroma_width;
       cu->type = bottom_right->type;
       cu->tr_skip |= bottom_right->tr_skip & 6;
     }
   }
 }
-  
 
 
+static void lcu_fill_chroma_cbfs(lcu_t *lcu, const cu_loc_t * const chroma_loc, enum uvg_tree_type tree_type)
+{
+  int8_t height = tree_type == UVG_CHROMA_T ? chroma_loc->chroma_height : chroma_loc->height;
+  int8_t width = tree_type == UVG_CHROMA_T ? chroma_loc->chroma_width : chroma_loc->width;
+  uint32_t x_local = chroma_loc->local_x;
+  uint32_t y_local = chroma_loc->local_y;
+  const int offset = ~((TR_MAX_WIDTH >> (tree_type == UVG_CHROMA_T)) - 1);
+  // Set coeff flags in every CU covered by part_mode in this depth.
+  for (uint32_t y = 0; y < height; y += SCU_WIDTH) {
+    for (uint32_t x = 0; x < width; x += SCU_WIDTH) {
+      // Use TU top-left CU to propagate coeff flags
+      cu_info_t* cu_from = LCU_GET_CU_AT_PX(lcu, x_local + (x & offset), y_local + (y & offset));
+      cu_info_t* cu_to = LCU_GET_CU_AT_PX(lcu, x_local + x, y_local + y);
+      if (cu_from != cu_to) {
+        cbf_copy(&cu_to->cbf, cu_from->cbf, COLOR_U);
+        cbf_copy(&cu_to->cbf, cu_from->cbf, COLOR_V);
+      }
+    }
+  }
+  
+}
+
 static void lcu_fill_cbf(lcu_t *lcu, int x_local, unsigned y_local, unsigned width, unsigned height, const cu_info_t *cur_cu, enum
                          uvg_tree_type tree_type)
 {
@@ -996,6 +1027,97 @@ void uvg_sort_keys_by_cost(unit_stats_map_t *__restrict map)
 }
 
 
+static void mark_deblocking(const cu_loc_t* const cu_loc, const cu_loc_t* const chroma_loc, lcu_t* lcu, enum uvg_tree_type tree_type, bool has_chroma, const bool is_separate_tree, int x_local, int y_local)
+{
+  if(tree_type != UVG_CHROMA_T) {
+    if(cu_loc->x) {
+      for (int x = cu_loc->local_x; x < cu_loc->local_x + cu_loc->width; x += TR_MAX_WIDTH) {
+        for (int y = cu_loc->local_y; y < cu_loc->local_y + cu_loc->height; y += SCU_WIDTH) {
+          LCU_GET_CU_AT_PX(lcu, x, y)->luma_deblocking |= EDGE_VER;
+          if(!is_separate_tree && tree_type == UVG_BOTH_T) LCU_GET_CU_AT_PX(lcu, x, y)->chroma_deblocking |= EDGE_VER;
+        }
+      }
+    }
+    else if(cu_loc->width == 64) {
+      for (int y = cu_loc->local_y; y < cu_loc->local_y + cu_loc->height; y += SCU_WIDTH) {
+        LCU_GET_CU_AT_PX(lcu, TR_MAX_WIDTH, y)->luma_deblocking |= EDGE_VER;
+        if (!is_separate_tree && tree_type == UVG_BOTH_T) LCU_GET_CU_AT_PX(lcu, TR_MAX_WIDTH, y)->chroma_deblocking |= EDGE_VER;
+      }        
+    }
+
+    if(cu_loc->y) {
+      for (int y = cu_loc->local_y; y < cu_loc->local_y + cu_loc->height; y += TR_MAX_WIDTH) {
+        for (int x = cu_loc->local_x; x < cu_loc->local_x + cu_loc->width; x += SCU_WIDTH) {
+          LCU_GET_CU_AT_PX(lcu, x, y)->luma_deblocking |= EDGE_HOR;
+          if (!is_separate_tree && tree_type == UVG_BOTH_T) LCU_GET_CU_AT_PX(lcu, x, y)->chroma_deblocking |= EDGE_HOR;
+        }
+      }
+    }
+    else if (cu_loc->height == 64) {
+      for (int x = cu_loc->local_x; x < cu_loc->local_x + cu_loc->width; x += SCU_WIDTH) {
+        LCU_GET_CU_AT_PX(lcu, x, TR_MAX_WIDTH)->luma_deblocking |= EDGE_VER;
+        if (!is_separate_tree && tree_type == UVG_BOTH_T) LCU_GET_CU_AT_PX(lcu, x, TR_MAX_WIDTH)->chroma_deblocking |= EDGE_VER;
+      }
+    }
+
+    if(is_separate_tree && has_chroma) {
+      if (chroma_loc->x) {
+        for (int x = chroma_loc->local_x; x < chroma_loc->local_x + chroma_loc->width; x += TR_MAX_WIDTH) {
+          for (int y = chroma_loc->local_y; y < chroma_loc->local_y + chroma_loc->height; y += SCU_WIDTH) {
+            LCU_GET_CU_AT_PX(lcu, x, y)->chroma_deblocking |= EDGE_VER;
+          }
+        }
+      }
+      else if(cu_loc->width == 64) {
+        for (int y = chroma_loc->local_y; y < chroma_loc->local_y + chroma_loc->height; y += SCU_WIDTH) {
+          LCU_GET_CU_AT_PX(lcu, TR_MAX_WIDTH, y)->chroma_deblocking |= EDGE_VER;
+        }          
+      }
+
+      if (chroma_loc->y) {
+        for (int y = chroma_loc->local_y; y < chroma_loc->local_y + chroma_loc->height; y += TR_MAX_WIDTH) {
+          for (int x = chroma_loc->local_x; x < chroma_loc->local_x + chroma_loc->width; x += SCU_WIDTH) {
+            LCU_GET_CU_AT_PX(lcu, x, y)->chroma_deblocking |= EDGE_HOR;
+          }
+        }
+      }
+      else if (cu_loc->height == 64) {
+        for (int x = chroma_loc->local_x; x < chroma_loc->local_x + chroma_loc->width; x += SCU_WIDTH) {
+          LCU_GET_CU_AT_PX(lcu, x, TR_MAX_WIDTH)->chroma_deblocking |= EDGE_VER;
+        }
+      }
+    }
+  }
+  else {
+
+    if (chroma_loc->x) {
+      for (int x = x_local; x < x_local + chroma_loc->chroma_width; x += TR_MAX_WIDTH / 2) {
+        for (int y = y_local; y < y_local + chroma_loc->chroma_height; y += SCU_WIDTH) {
+          LCU_GET_CU_AT_PX(lcu, x, y)->chroma_deblocking |= EDGE_VER;
+        }
+      }
+    }
+    else if(chroma_loc->width == 64) {
+      for (int y = y_local; y < y_local + chroma_loc->chroma_height; y += SCU_WIDTH) {
+        LCU_GET_CU_AT_PX(lcu, TR_MAX_WIDTH / 2, y)->chroma_deblocking |= EDGE_VER;
+      }        
+    }
+
+    if(chroma_loc->y) {
+      for (int y = y_local; y < y_local + chroma_loc->chroma_height; y += TR_MAX_WIDTH / 2) {
+        for (int x = x_local; x < x_local + chroma_loc->chroma_width; x += SCU_WIDTH) {
+          LCU_GET_CU_AT_PX(lcu, x, y)->chroma_deblocking |= EDGE_HOR;
+        }
+      }        
+    }
+    else if (chroma_loc->height == 64) {
+      for (int x = x_local; x < x_local + chroma_loc->chroma_width; x += SCU_WIDTH) {
+        LCU_GET_CU_AT_PX(lcu, x, TR_MAX_WIDTH / 2)->chroma_deblocking |= EDGE_VER;
+      }
+    }
+  }
+}
+
 /**
  * Search every mode from 0 to MAX_PU_DEPTH and return cost of best mode.
  * - The recursion is started at depth 0 and goes in Z-order to MAX_PU_DEPTH.
@@ -1090,6 +1212,11 @@ static double search_cu(
   cur_cu->log2_width = uvg_g_convert_to_log2[cu_width];
   cur_cu->log2_height = uvg_g_convert_to_log2[cu_height];
 
+  if(chroma_loc) {
+    cur_cu->log2_chroma_height = uvg_g_convert_to_log2[chroma_loc->chroma_height];
+    cur_cu->log2_chroma_width = uvg_g_convert_to_log2[chroma_loc->chroma_width];
+  }
+
   // If the CU is completely inside the frame at this depth, search for
   // prediction modes at this depth.
   if ( x + luma_width <= frame_width && y + luma_height <= frame_height)
@@ -1269,23 +1396,21 @@ static double search_cu(
       if((!recon_chroma && state->encoder_control->chroma_format != UVG_CSP_400 && tree_type != UVG_LUMA_T) 
         || tree_type == UVG_CHROMA_T) {
         intra_search.pred_cu.intra.mode_chroma = cur_cu->intra.mode_chroma;
-        lcu_fill_chroma_cu_info(
-          lcu,
-          chroma_loc);
+        if(tree_type != UVG_CHROMA_T) {
+          lcu_fill_chroma_cu_info(
+            lcu,
+            chroma_loc);
+        }
         uvg_intra_recon_cu(state,
                            &intra_search, chroma_loc,
                            NULL, lcu,
                            UVG_CHROMA_T,
                            false,
                            true);
-        lcu_fill_cbf(
+        lcu_fill_chroma_cbfs(
           lcu,
-          chroma_loc->local_x,
-          chroma_loc->local_y,
-          chroma_loc->width,
-          chroma_loc->height,
-          cur_cu,
-          UVG_CHROMA_T);
+          chroma_loc,
+          tree_type);
       } else {
         assert(cur_cu->cr_lfnst_idx == 0 && "If we don't have separate tree chroma lfnst index must be 0");
       }
@@ -1409,6 +1534,16 @@ static double search_cu(
     //  lcu_fill_cbf(lcu, x_local, y_local, cu_width, cur_cu);
     //}
     cabac->update = 0;
+
+    mark_deblocking(
+      cu_loc,
+      chroma_loc,
+      lcu,
+      tree_type,
+      has_chroma,
+      is_separate_tree,
+      x_local,
+      y_local);
   } 
 
   bool can_split_cu =