diff --git a/conditioning.py b/conditioning.py
index b7a20078..2d4a6cdb 100644
--- a/conditioning.py
+++ b/conditioning.py
@@ -9920,7 +9920,7 @@ class FunPackVideoRefinerV2(FunPackVideoRefiner):
                   prompt_enhance_temperature=0.7, prompt_enhance_top_p=0.92,
                   prompt_enhance_max_length=400, prompt_enhance_thinking=False,
                   prompt_enhance_image=None, prompt_enhance_sampling=None,
-                  prompt_enhance_output=False,
+                  prompt_enhance_output=False, prompt_enhance_scenes=True,
                   _seed=None, _seed_source="fresh seed", _scene_seeds=None, _velocity_keys=None,
                   batch_variants=1, guess_mode=False, guess_direction="up", guess_range=1.0,
                   guess_freeze_seed=True, movie_editor_scene_ratings=None, scene_segments=None,
@@ -10571,11 +10571,17 @@ class FunPackVideoRefinerV2(FunPackVideoRefiner):
         # When the run splits into scenes, the per-scene conditionings REPLACE this base entry
         # — so enhancing the base too would be a second full generation whose result is then
         # thrown away. One generation per prompt that is actually encoded, never two.
-        _enhance_base = prompt_enhance and not split_by_transitions
+        # Off when nothing reads Studio's own conditioning (the Editor checks the wiring):
+        # rewriting a prompt whose encoding is discarded is a generation for nothing.
+        _enhance_own = prompt_enhance and prompt_enhance_scenes
+        if prompt_enhance and not prompt_enhance_scenes:
+            print("[FunPackVideoRefinerV2] Prompt enhancer: Studio's own prompt skipped — "
+                  "nothing reads its conditioning; only the enhanced_prompt output is written.")
+        _enhance_base = _enhance_own and not split_by_transitions
         # top_k / min_p / penalties / greedy; `seed` 0 = follow the generation seed.
         _enhance_sampling = dict(prompt_enhance_sampling or {})
         _enhance_seed = int(_enhance_sampling.pop("seed", 0) or 0) or seed
-        if prompt_enhance and not _enhance_base:
+        if _enhance_own and not _enhance_base:
             print("[FunPackVideoRefinerV2] Prompt enhancer: base prompt skipped — this run "
                   "splits into scenes and each scene is enhanced instead.")
         if _enhance_base:
@@ -10845,7 +10851,7 @@ class FunPackVideoRefinerV2(FunPackVideoRefiner):
                     # authoritative on count, so rewriting the scenes as a single paragraph
                     # and re-splitting could return a different number of scenes and desync
                     # every clip from its anchor. One call per distinct text, cached.
-                    if prompt_enhance and split_scene_texts:
+                    if _enhance_own and split_scene_texts:
                         _enhanced_texts = []
                         for _i, t in enumerate(split_scene_texts):
                             _after, _st = self._v2_enhance_prompt(
@@ -13845,6 +13851,7 @@ class FunPackStudio:
             h3_phrase_variability=float(rf.get("h3_phrase_variability", 0.0) or 0.0),
             prompt_enhance=bool(rf.get("prompt_enhance", False)),
             prompt_enhance_output=bool(rf.get("prompt_enhance_output", False)),
+            prompt_enhance_scenes=bool(rf.get("prompt_enhance_scenes", True)),
             prompt_enhance_system=str(rf.get("prompt_enhance_system", "") or ""),
             prompt_enhance_temperature=float(rf.get("prompt_enhance_temperature", 0.7)),
             prompt_enhance_top_p=float(rf.get("prompt_enhance_top_p", 0.92)),
diff --git a/movie_editor/backend/builder.py b/movie_editor/backend/builder.py
index ed0b1273..4cb6acc1 100644
--- a/movie_editor/backend/builder.py
+++ b/movie_editor/backend/builder.py
@@ -1052,6 +1052,20 @@ def build(object_info: dict, models_config: dict, params: dict, media: dict | No
     _break_cycles(graph, report, protected_edges,
                   _node_labels(slots, slot_node_id, object_info, core=CORE))
 
+    # Nothing reads Studio's own conditioning (e.g. H3's reference-to-video node feeding the
+    # sampler directly): Studio's per-scene prompt enhancement would be thrown away.
+    if "studio" in graph and "sampler" in graph and \
+            not _feeds_from(graph, graph["sampler"]["inputs"].get("positive"), ("studio", 1)):
+        try:
+            _ss_p = json.loads(str(graph["studio"]["inputs"].get("studio_settings") or "{}"))
+        except ValueError:
+            _ss_p = {}
+        if isinstance(_ss_p, dict):
+            _rf_p = _ss_p.get("refiner") if isinstance(_ss_p.get("refiner"), dict) else {}
+            _rf_p["prompt_enhance_scenes"] = False
+            _ss_p["refiner"] = _rf_p
+            graph["studio"]["inputs"]["studio_settings"] = json.dumps(_ss_p)
+
     for msg in pipeline_wiring.validate_models_wiring(models_config):
         report["blocking"].append(msg)
         report["unsatisfied"].append(msg)
@@ -1373,6 +1387,22 @@ def _reaches_upstream(graph, start, goal):
     return False
 
 
+def _feeds_from(graph, value, source):
+    """Whether the link `value` is, or is built upstream from, output `source` (node, idx)."""
+    stack, seen = [value], set()
+    while stack:
+        v = stack.pop()
+        if not (isinstance(v, list) and len(v) == 2 and isinstance(v[0], str)):
+            continue
+        if (v[0], v[1]) == tuple(source):
+            return True
+        if v[0] in seen or v[0] not in graph:
+            continue
+        seen.add(v[0])
+        stack.extend((graph[v[0]].get("inputs") or {}).values())
+    return False
+
+
 def _edge_input(graph, dst, up):
     """The input name on `dst` that consumes `up`'s output (or None)."""
     for inp, val in (graph.get(dst, {}).get("inputs", {}) or {}).items():
diff --git a/movie_editor/tests/test_builder.py b/movie_editor/tests/test_builder.py
index 27d92031..85f5cf00 100644
--- a/movie_editor/tests/test_builder.py
+++ b/movie_editor/tests/test_builder.py
@@ -1597,3 +1597,19 @@ def test_enhanced_prompt_link_wires_to_studios_output():
     assert oi["FunPackStudio"]["output_name"][builder.STUDIO_ENHANCED_PROMPT_OUT] == "enhanced_prompt"
     rf = json.loads(graph["studio"]["inputs"]["studio_settings"])["refiner"]
     assert rf["prompt_enhance_output"] is True
+
+
+def test_studio_skips_its_own_enhancement_when_nothing_reads_its_conditioning():
+    import json
+    oi = {**OI, "R2V": {"input": {"required": {"prompt": ["STRING", {"default": ""}]}},
+                        "output": ["CONDITIONING"], "output_name": ["CONDITIONING"]}}
+
+    def rf_of(models):
+        graph, _ = builder.build(oi, models, PARAMS)
+        return json.loads(graph["studio"]["inputs"]["studio_settings"]).get("refiner", {})
+
+    assert "prompt_enhance_scenes" not in rf_of({"slots": []})      # default: Studio feeds it
+    bypassed = {"full_control": True,
+                "slots": [{"id": "r2v", "node_class": "R2V", "inputs": {}, "wires": {}}],
+                "core_overrides": {"sampler": {"positive": "out:r2v:CONDITIONING"}}}
+    assert rf_of(bypassed)["prompt_enhance_scenes"] is False
diff --git a/tests/test_prompt_enhancer.py b/tests/test_prompt_enhancer.py
index 3e8b8832..a88f860d 100644
--- a/tests/test_prompt_enhancer.py
+++ b/tests/test_prompt_enhancer.py
@@ -345,7 +345,7 @@ def _refine_source():
 
 def test_the_base_prompt_is_not_enhanced_when_the_run_splits():
     src = _refine_source()
-    assert "_enhance_base = prompt_enhance and not split_by_transitions" in src
+    assert "_enhance_base = _enhance_own and not split_by_transitions" in src
 
 
 def test_the_base_enhancement_is_gated_on_that_flag():
@@ -356,7 +356,7 @@ def test_the_base_enhancement_is_gated_on_that_flag():
 
 def test_the_scene_path_is_still_enhanced_when_splitting():
     src = _refine_source()
-    assert "if prompt_enhance and split_scene_texts:" in src
+    assert "if _enhance_own and split_scene_texts:" in src
 
 
 def test_the_skip_is_reported_not_silent():
diff --git a/tests/test_v2_category_alignment.py b/tests/test_v2_category_alignment.py
index 48df95f3..e09028f8 100644
--- a/tests/test_v2_category_alignment.py
+++ b/tests/test_v2_category_alignment.py
@@ -2321,3 +2321,11 @@ def test_enhanced_prompt_output_costs_nothing_when_unlinked(tmp_path):
     out, clip = _enhanced_output(tmp_path, prompt_enhance=True)
     assert out is None
     assert not any(t.endswith("a cat, soft light") for t, _ in clip.tokenize_calls)
+
+
+def test_studio_skips_its_own_enhancement_when_told_nothing_reads_it(tmp_path):
+    out, clip = _enhanced_output(tmp_path, prompt_enhance=True, prompt_enhance_scenes=False)
+    assert out is None and clip.generate_kwargs == {}
+    out, clip = _enhanced_output(tmp_path, prompt_enhance=True, prompt_enhance_scenes=False,
+                                 prompt_enhance_output=True)
+    assert out == "a vivid cat on a sunlit sill"