diff --git a/ltx_enhancements.py b/ltx_enhancements.py
index 9adcab83..b7ad7770 100644
--- a/ltx_enhancements.py
+++ b/ltx_enhancements.py
@@ -566,6 +566,18 @@ def build_enhancements(model, rating_profile, temporal_style, refinement_key, re
     }
 
     # --- Technique 5: injection data ---
+    # Clear the temp maps file so this generation's capture starts fresh.
+    # Without this, maps from previous runs merge in and compound across
+    # sessions — each "loved it" cycle injects tainted states from the
+    # previous round, producing progressively worse color bias.
+    if refinement_key:
+        temp_path = _temp_maps_path(refinement_key)
+        try:
+            if os.path.exists(temp_path):
+                os.remove(temp_path)
+        except Exception:
+            pass
+
     has_i2v = reference_latent is not None and _has_i2v_reference(reference_latent)
     if has_i2v:
         # i2v latent: extract reference maps on first generation step (real conditioning available).
diff --git a/samplers.py b/samplers.py
index d52ce8c5..59ffe4c4 100644
--- a/samplers.py
+++ b/samplers.py
@@ -1467,7 +1467,10 @@ class FunPackLTXAVSceneChainSampler:
                 # Record boundary (center of the overlap region) before blending
                 effect = self._scene_transition_effect(scene_cond)
                 if effect and transition_duration > 0:
-                    boundary_latent = max(0, cumulative_latent_frames - video_overlap // 2)
+                    # Seam is at cumulative_latent_frames: overlap frames are pinned
+                    # (mask=0) so the latent blend is a no-op — the actual content
+                    # boundary is exactly here, not offset by overlap//2.
+                    boundary_latent = cumulative_latent_frames
                     boundary_pixel = int((boundary_latent - 1) * time_scale + 1) if time_scale > 1 else boundary_latent
                     blend_half_latent = max(1, video_overlap // 2)
                     blend_half_pixels = blend_half_latent * time_scale