diff --git a/shot_memory.py b/shot_memory.py
index bf57d868..ec60deb8 100644
--- a/shot_memory.py
+++ b/shot_memory.py
@@ -225,6 +225,10 @@ class ShotMemory:
         coarse = own
         if parent is not None:
             coarse = amount * _fit(parent["coarse"], *own.shape[-2:]) + math.sqrt(1.0 - amount ** 2) * own
+            # Unit variance only holds when the two layouts are unrelated. Rerunning the
+            # liked shot's own seed makes them the same pattern, and the blend stacks it
+            # (1.41x at amount 0.7): back to a normal-strength layout, per channel.
+            coarse = coarse / coarse.std(dim=(1, 2), keepdim=True).clamp(min=1e-6)
             video = with_coarse(video, coarse)
         if record and not self.used:
             self.used.append({
diff --git a/tests/test_shot_memory_tsr.py b/tests/test_shot_memory_tsr.py
index 9ea513b9..2f49cae6 100644
--- a/tests/test_shot_memory_tsr.py
+++ b/tests/test_shot_memory_tsr.py
@@ -205,3 +205,14 @@ def test_one_rating_is_one_row_and_skips_are_counted(monkeypatch, tmp_path):
     assert len(mem.used) == 1 and mem.skipped == 1
     mem.save_pending()
     assert sm.commit("k", 1.0) == 1
+
+
+def test_reusing_a_shot_with_its_own_seed_keeps_normal_strength(monkeypatch, tmp_path):
+    """Same seed as the liked shot: the two layouts are one pattern, and a plain blend
+    would stack it (1.41x at amount 0.7)."""
+    _store(monkeypatch, tmp_path)
+    noise = torch.randn(1, 24, 7, 24, 36, generator=torch.Generator().manual_seed(5))
+    mem = sm.ShotMemory("k", "manual", 0.7, rng=random.Random(0))
+    mem.plans[(None, 24)] = ({"id": 1, "coarse": sm.coarse_of(noise).half()}, 0.7)
+    out = mem.shape(noise.clone(), torch.zeros_like(noise), None)
+    assert abs(float(sm.coarse_of(out).std()) - 1.0) < 0.02