Naturalize caption character expressions

caption_metadata_routes.py

@@ -94,6 +94,26 @@ def couple_subject_sentence(
     return f"{subject} are adults"
 
 
+def expression_detail(expression: Any, clean_text: Callable[[Any], str]) -> tuple[str, bool]:
+    text = clean_text(expression)
+    if not text:
+        return "", False
+    has_character_labels = bool(
+        re.search(
+            r"\b(?:Woman|Man) [A-Z] has\b|\bthe (?:woman|man) has\b",
+            text,
+            flags=re.IGNORECASE,
+        )
+    )
+    text = re.sub(
+        r"\b((?:Woman|Man) [A-Z]|the (?:woman|man)) has\b",
+        r"\1 with",
+        text,
+        flags=re.IGNORECASE,
+    )
+    return text, has_character_labels
+
+
 def single_from_row_result(
     request: CaptionMetadataRouteRequest,
     deps: CaptionMetadataRouteDependencies,
@@ -148,7 +168,11 @@ def single_from_row_result(
     if pose:
         parts.append(f"{pronoun(subject)} is {deps.pose_clause(pose)}")
     if expression:
-        parts.append(f"{possessive_pronoun(subject)} expression is {expression}")
+        expression, labeled_expression = expression_detail(expression, deps.clean_text)
+        if labeled_expression:
+            parts.append(f"The expression detail shows {expression}")
+        else:
+            parts.append(f"{possessive_pronoun(subject)} expression is {expression}")
     if scene:
         parts.append(f"The setting is {scene}")
     if deps.detail_allows(detail_level) and camera_scene:
@@ -204,7 +228,11 @@ def couple_from_row_result(
     if deps.detail_allows(detail_level) and camera_scene:
         parts.append(camera_scene)
     if expression:
-        parts.append(f"Their expressions are {expression}")
+        expression, labeled_expression = expression_detail(expression, deps.clean_text)
+        if labeled_expression:
+            parts.append(f"The expression details show {expression}")
+        else:
+            parts.append(f"Their expressions are {expression}")
     if deps.detail_allows(detail_level) and composition:
         parts.append(f"The composition is {composition}")
     if keep_style and style:
@@ -258,7 +286,11 @@ def configured_cast_from_row_result(
     if scene:
         scene_bits.append(f"set in {scene}")
     if expression:
-        scene_bits.append(f"with {expression}")
+        expression, labeled_expression = expression_detail(expression, deps.clean_text)
+        if labeled_expression:
+            scene_bits.append(f"with expression details showing {expression}")
+        else:
+            scene_bits.append(f"with {expression}")
     if composition:
         scene_bits.append(f"framed as {composition}")
     if scene_bits and deps.detail_allows(detail_level):
@@ -299,7 +331,11 @@ def group_or_layout_from_row_result(
     if primary == "layout scene":
         parts = [f"{deps.cap_first(subject)} is arranged as an adults-only designed illustration layout"]
         if expression:
-            parts.append(f"The featured expression is {expression}")
+            expression, labeled_expression = expression_detail(expression, deps.clean_text)
+            if labeled_expression:
+                parts.append(f"The featured expression details show {expression}")
+            else:
+                parts.append(f"The featured expression is {expression}")
     else:
         parts = [f"{deps.cap_first(subject)} includes adults"]
         if age:
@@ -307,7 +343,11 @@ def group_or_layout_from_row_result(
         if item:
             parts.append(f"They wear {item}")
         if expression:
-            parts.append(f"They show {expression}")
+            expression, labeled_expression = expression_detail(expression, deps.clean_text)
+            if labeled_expression:
+                parts.append(f"Their expressions show {expression}")
+            else:
+                parts.append(f"They show {expression}")
     if scene:
         parts.append(f"The setting is {scene}")
     if deps.detail_allows(detail_level) and camera_scene:

tools/prompt_route_simulation.py

@@ -526,6 +526,16 @@ def _caption_cast_descriptor_issues(name: str, row: dict[str, Any] | None, capti
     return []
 
 
+def _caption_expression_grammar_issues(name: str, caption_text: str) -> list[str]:
+    if re.search(
+        r"\b(?:with|are|show|shows|is|include|includes)\s+(?:woman|man) [a-z]\s+has\b",
+        caption_text,
+        flags=re.IGNORECASE,
+    ):
+        return [f"{name}.caption: character_expression_has_grammar"]
+    return []
+
+
 def _trace_dict(formatter_name: str, payload: dict[str, Any]) -> tuple[dict[str, Any], str]:
     trace_text = str(payload.get("route_trace_json") or "")
     if not trace_text:
@@ -659,6 +669,7 @@ def _formatter_issues(
             issues.append(f"{name}.{formatter_name}: not_metadata_route:{method}")
     issues.extend(_formatter_trace_issues(name, formats, target=target, row=row))
     issues.extend(_caption_cast_descriptor_issues(name, row, caption_text))
+    issues.extend(_caption_expression_grammar_issues(name, caption_text))
 
     for label, value in (
         (f"{name}.krea_negative", krea.get("negative_prompt")),

tools/prompt_smoke.py

@@ -4262,6 +4262,27 @@ def smoke_caption_metadata_routes() -> None:
         caption_naturalizer._configured_cast_from_row,
         "metadata(configured_cast)",
     )
+    configured_character_expression = _fixture_hardcore_row(
+        character_expression_text="Woman A has flushed focus; Man A has concentrated stare",
+    )
+    character_expression_route = caption_metadata_routes.configured_cast_from_row_result(
+        caption_naturalizer._caption_metadata_route_request(configured_character_expression, "balanced", False),
+        caption_naturalizer._caption_metadata_route_dependencies(),
+    )
+    _expect(character_expression_route is not None, "Caption configured-cast character expression row did not match")
+    assert character_expression_route is not None
+    _expect(
+        "with Woman A has" not in character_expression_route.prose,
+        "Caption configured-cast prose kept old character-expression grammar",
+    )
+    _expect(
+        "Woman A with flushed focus" in character_expression_route.prose,
+        "Caption configured-cast prose did not naturalize Woman A expression",
+    )
+    _expect(
+        "Man A with concentrated stare" in character_expression_route.prose,
+        "Caption configured-cast prose did not naturalize Man A expression",
+    )
     configured_axis_only = _fixture_hardcore_row(
         item="generic configured adult action",
         role_graph="",