Share fallback field-label cleanup

docs/prompt-architecture-improvement-plan.md

@@ -75,6 +75,8 @@ routes:
 - `Avoid:` positive/negative splitting for fallback text;
 - the shared prompt field-label inventory and extraction such as `Setting:`,
   `Sexual scene:`, `Camera control:`, or `Composition:`;
+- fallback field-label stripping for tag/text routes that need label-free body
+  text;
 - row-value fallback from metadata fields to labeled prompt text.
 
 It must not make formatter-style decisions. Krea prose, SDXL tags, and training
@@ -291,10 +293,11 @@ Keep here:
 - formatter profiles for manual controls, Pony flat-vector, SDXL photo, and
   plain flat-vector styles live in `sdxl_presets.py` and are exposed by
   `SxCP SDXL Formatter`.
+- fallback field-label cleanup delegates to `formatter_input.py`.
 
 Improve later:
-
+- add route-level fixtures for any new SDXL model profile that needs different
-- make fallback cleanup use the shared field-label inventory.
+  tag ordering.
 
 ### Naturalizer Path
 

docs/prompt-pool-routing-map.md

@@ -95,7 +95,7 @@ Core helper ownership:
 | `krea_cast.py` | Shared formatter cast descriptor parsing, cast labels, cast prose, natural cast descriptor text, and label replacement used by Krea2 and caption routes. |
 | `prompt_hygiene.py` | Generic prompt, caption, and negative-prompt cleanup. |
 | `row_normalization.py` | Final prompt-row and pair metadata normalization: trigger prepending, extra-positive append, negative merge/dedupe, caption-part joining, and embedded soft/hard row sanitation. |
-| `formatter_input.py` | Shared formatter input parsing: text cleanup, metadata/source JSON detection, trigger-prefix stripping, shared prompt field-label inventory, `Avoid:` splitting, prompt-field extraction, and metadata row-value fallback. |
+| `formatter_input.py` | Shared formatter input parsing: text cleanup, metadata/source JSON detection, trigger-prefix stripping, shared prompt field-label inventory, fallback field-label stripping, `Avoid:` splitting, prompt-field extraction, and metadata row-value fallback. |
 | `sdxl_presets.py` | SDXL formatter profiles, style presets, quality presets, default negative prompt, and metadata-family tag hints used by the SDXL formatter and node choice lists. |
 | `caption_policy.py` | Caption naturalizer policy data and helpers: caption profiles, style tails, item labels, metadata-family caption labels, detail/style-policy normalization, clothing cleanup, and composition cleanup. |
 
@@ -661,7 +661,7 @@ SDXL field consumption:
 | Normal metadata | cast descriptors, age/body/skin/hair/eyes, `action_family`, `position_family`, `position_keys`, item, role graph, scene, camera config/directive | `_row_core_tags`, `_metadata_family_tags`, `_camera_tags` |
 | Pair softcore | `softcore_row`, pair partner styling, root soft camera config | `_soft_tags` |
 | Pair hardcore | `hardcore_row`, `action_family`, `position_family`, `position_keys`, `hardcore_clothing_state`, hard camera fields, hard prompt text | `_hard_tags`, `_metadata_family_tags` |
-| Text fallback | `source_text`, preserve-trigger setting | `_fallback_text_to_sdxl` |
+| Text fallback | `source_text`, preserve-trigger setting, shared field-label stripping | `_fallback_text_to_sdxl` |
 
 SDXL is the right place for model trigger handling, tag ordering, weight syntax,
 quality/style preset changes, and nude-weight defaults. Do not solve those in

formatter_input.py

@@ -101,6 +101,18 @@ def split_avoid(text: Any) -> tuple[str, str]:
     return text[: match.start()].strip(" ."), match.group(1).strip(" .")
 
 
+def strip_prompt_field_labels(
+    text: Any,
+    *,
+    field_labels: tuple[str, ...] | list[str] = DEFAULT_PROMPT_FIELD_LABELS,
+) -> str:
+    text = clean_text(text)
+    if not text:
+        return ""
+    labels = "|".join(re.escape(name) for name in sorted(field_labels, key=len, reverse=True))
+    return clean_text(re.sub(rf"\b(?:{labels}):\s*", "", text))
+
+
 def prompt_field(
     text: Any,
     label: str,

sdxl_formatter.py

@@ -63,6 +63,10 @@ def _split_avoid(text: str) -> tuple[str, str]:
     return input_policy.split_avoid(text)
 
 
+def _strip_prompt_field_labels(text: str) -> str:
+    return input_policy.strip_prompt_field_labels(text, field_labels=PROMPT_FIELD_LABELS)
+
+
 def _prompt_field(text: str, label: str) -> str:
     return input_policy.prompt_field(text, label, field_labels=PROMPT_FIELD_LABELS)
 
@@ -416,11 +420,7 @@ def _fallback_text_to_sdxl(
     nude_weight: float,
 ) -> tuple[str, str, str]:
     positive, negative = _split_avoid(_strip_trigger(source_text, preserve_trigger))
-    positive = re.sub(
+    positive = _strip_prompt_field_labels(positive)
-        r"\b(?:Scene|Setting|Pose|Sexual pose|Sexual scene|Facial expressions?|Composition|Role graph|Camera control):\s*",
-        "",
-        positive,
-    )
     tags = _combine_tags(positive, ", ".join(_explicit_tags(positive, nude_weight)))
     return tags, negative, "text(fallback)"
 

tools/prompt_smoke.py

@@ -904,9 +904,19 @@ def smoke_formatter_input_policy() -> None:
         formatter_input.prompt_field(labeled_prompt, "Camera control") == "side view",
         "Shared formatter field-label inventory lost Camera control parsing",
     )
+    stripped_labels = formatter_input.strip_prompt_field_labels(
+        "Characters: woman. Erotic outfit: sheer dress. Camera: side view."
+    )
+    _expect("Characters:" not in stripped_labels, "Shared label stripper did not remove Characters label")
+    _expect("Erotic outfit:" not in stripped_labels, "Shared label stripper did not remove Erotic outfit label")
+    _expect("Camera:" not in stripped_labels, "Shared label stripper did not remove Camera label")
 
     _expect(krea_formatter._clean("a  b , c") == formatter_input.clean_text("a  b , c"), "Krea clean helper is not delegated")
     _expect(sdxl_formatter._clean("a  b , c") == formatter_input.clean_text("a  b , c"), "SDXL clean helper is not delegated")
+    _expect(
+        sdxl_formatter._strip_prompt_field_labels("Characters: woman. Camera: side view.") == "woman. side view.",
+        "SDXL label stripper should delegate to formatter_input",
+    )
     _expect(caption_naturalizer._clean_text("a  b , c") == formatter_input.clean_text("a  b , c"), "Caption clean helper is not delegated")
     _expect(krea_formatter._strip_trigger(f"{Trigger}, prompt text", False) == "prompt text", "Krea trigger stripping changed")
     _expect(sdxl_formatter._strip_trigger(f"{SdxlTrigger}, prompt text", False) == "prompt text", "SDXL trigger stripping changed")
@@ -921,6 +931,19 @@ def smoke_formatter_input_policy() -> None:
     _expect_text("formatter_input.krea_prompt", krea.get("krea_prompt"), 20)
     _expect_text("formatter_input.sdxl_prompt", sdxl.get("sdxl_prompt"), 20)
     _expect_text("formatter_input.caption", caption, 20)
+    fallback_sdxl = sdxl_formatter.format_sdxl_prompt(
+        "Characters: woman. Erotic outfit: sheer dress. Camera: side view. Avoid: blur",
+        input_hint="prompt",
+        style_preset="none",
+        quality_preset="none",
+        trigger=SdxlTrigger,
+        prepend_trigger=False,
+    )
+    fallback_prompt = fallback_sdxl.get("sdxl_prompt", "")
+    _expect("Characters:" not in fallback_prompt, "SDXL fallback leaked Characters label")
+    _expect("Erotic outfit:" not in fallback_prompt, "SDXL fallback leaked Erotic outfit label")
+    _expect("Camera:" not in fallback_prompt, "SDXL fallback leaked Camera label")
+    _expect("blur" in fallback_sdxl.get("negative_prompt", ""), "SDXL fallback lost Avoid negative text")
 
 
 def smoke_formatter_cast_policy() -> None: