diff --git a/av1/encoder/partition_search.c b/av1/encoder/partition_search.c
index a41cc0c..38f423a 100644
--- a/av1/encoder/partition_search.c
+++ b/av1/encoder/partition_search.c
@@ -5373,27 +5373,41 @@
   start_timing(cpi, none_partition_search_time);
 #endif
 
-  // Further pruning or in some cases reverse pruning when allintra is set
-  // This code helps visual and in some cases metrics quality where the current
-  // block comprises at least one very low variance sub-block and at least one
-  // where the variance is much higher.
-  //
-  // The idea is that in such cases there is danger of ringing and other visual
-  // artifacts from a high variance feature such as an edge into a very low
-  // variance region.
-  //
-  // The approach taken is to force break down / split to a smaller block size
-  // to try and separate out the low variance and well predicted blocks from the
-  // more complex ones and to prevent propagation of ringing over a large
-  // region.
-  if ((cpi->oxcf.mode == ALLINTRA) && (bsize >= BLOCK_16X16)) {
-    double var_min, var_max;
-    log_sub_block_var(cpi, x, bsize, &var_min, &var_max);
+  if (cpi->oxcf.mode == ALLINTRA) {
+    const bool bsize_at_least_16x16 = (bsize >= BLOCK_16X16);
+    const bool prune_rect_part_using_4x4_var_deviation =
+        (cpi->sf.part_sf.prune_rect_part_using_4x4_var_deviation &&
+         !x->must_find_valid_partition);
 
-    if ((var_min < 0.272) && ((var_max - var_min) > 3.0)) {
-      part_search_state.partition_none_allowed = 0;
-      part_search_state.terminate_partition_search = 0;
-      part_search_state.do_square_split = 1;
+    if (bsize_at_least_16x16 || prune_rect_part_using_4x4_var_deviation) {
+      double var_min, var_max;
+      log_sub_block_var(cpi, x, bsize, &var_min, &var_max);
+
+      // Further pruning or in some cases reverse pruning when allintra is set.
+      // This code helps visual and in some cases metrics quality where the
+      // current block comprises at least one very low variance sub-block and at
+      // least one where the variance is much higher.
+      //
+      // The idea is that in such cases there is danger of ringing and other
+      // visual artifacts from a high variance feature such as an edge into a
+      // very low variance region.
+      //
+      // The approach taken is to force break down / split to a smaller block
+      // size to try and separate out the low variance and well predicted blocks
+      // from the more complex ones and to prevent propagation of ringing over a
+      // large region.
+      if (bsize_at_least_16x16 && (var_min < 0.272) &&
+          ((var_max - var_min) > 3.0)) {
+        part_search_state.partition_none_allowed = 0;
+        part_search_state.terminate_partition_search = 0;
+        part_search_state.do_square_split = 1;
+      } else if (prune_rect_part_using_4x4_var_deviation &&
+                 (var_max - var_min < 3.0)) {
+        // Prune rectangular partitions if the variance deviation of 4x4
+        // sub-blocks within the block is less than a threshold (derived
+        // empirically).
+        part_search_state.do_rectangular_split = 0;
+      }
     }
   }
 
diff --git a/av1/encoder/speed_features.c b/av1/encoder/speed_features.c
index 7e244d0..9e42391 100644
--- a/av1/encoder/speed_features.c
+++ b/av1/encoder/speed_features.c
@@ -502,6 +502,7 @@
 
     sf->part_sf.prune_rectangular_split_based_on_qidx =
         allow_screen_content_tools ? 0 : 2;
+    sf->part_sf.prune_rect_part_using_4x4_var_deviation = true;
     sf->part_sf.prune_sub_8x8_partition_level =
         allow_screen_content_tools ? 0 : 1;
     sf->part_sf.prune_part4_search = 3;
@@ -1878,6 +1879,7 @@
   part_sf->rect_partition_eval_thresh = BLOCK_128X128;
   part_sf->prune_ext_part_using_split_info = 0;
   part_sf->prune_rectangular_split_based_on_qidx = 0;
+  part_sf->prune_rect_part_using_4x4_var_deviation = false;
   part_sf->early_term_after_none_split = 0;
   part_sf->ml_predict_breakout_level = 0;
   part_sf->prune_sub_8x8_partition_level = 0;
diff --git a/av1/encoder/speed_features.h b/av1/encoder/speed_features.h
index da0fd7a..4ff85f9 100644
--- a/av1/encoder/speed_features.h
+++ b/av1/encoder/speed_features.h
@@ -646,6 +646,18 @@
   // 2 : prune all block size based on qindex
   int prune_rectangular_split_based_on_qidx;
 
+  // Prune rectangular partitions based on 4x4 sub-block variance
+  // false : no pruning
+  // true : prune rectangular partitions based on 4x4 sub-block variance
+  // deviation
+  //
+  // For allintra encode, this speed feature reduces instruction count by 6.4%
+  // for speed=6 with coding performance change less than 0.24%. For AVIF image
+  // encode, this speed feature reduces encode time by 8.14% for speed 6 on a
+  // typical image dataset with coding performance change less than 0.16%. This
+  // speed feature is not applicable to speed >= 7.
+  bool prune_rect_part_using_4x4_var_deviation;
+
   // Terminate partition search for child partition,
   // when NONE and SPLIT partition rd_costs are INT64_MAX.
   int early_term_after_none_split;