From 7f6bf0ffd8efd4563980fa8de4571382948f4475 Mon Sep 17 00:00:00 2001
From: Ethanfel <ethan.fel@ts-pc.fr>
Date: Sat, 27 Jun 2026 22:06:37 +0200
Subject: [PATCH] Clean SDXL composition tags

---
 sdxl_tag_routes.py               | 11 +++++++++--
 tools/prompt_route_simulation.py |  8 ++++++++
 tools/prompt_smoke.py            |  9 +++++++++
 3 files changed, 26 insertions(+), 2 deletions(-)

diff --git a/sdxl_tag_routes.py b/sdxl_tag_routes.py
index f5621c7..f2608af 100644
--- a/sdxl_tag_routes.py
+++ b/sdxl_tag_routes.py
@@ -72,6 +72,13 @@ def _pair_counts(row: dict[str, Any], root: dict[str, Any]) -> tuple[int, int]:
     return _descriptor_counts(root)
 
 
+def _composition_tags_text(text: str) -> str:
+    text = re.sub(r"^vertical\s+", "", str(text or "").strip(), flags=re.IGNORECASE)
+    text = re.sub(r"\s+composition$", "", text, flags=re.IGNORECASE)
+    text = re.sub(r"\bcomposition\b", "frame", text, flags=re.IGNORECASE)
+    return text.strip(" ,")
+
+
 def _row_explicit_signal_text(
     row: dict[str, Any],
     *,
@@ -133,7 +140,7 @@ def row_core_tags_result(request: SDXLRowTagRequest, deps: SDXLTagRouteDependenc
         "expression",
         ("Facial expressions", "Facial expression"),
     )
-    composition = deps.row_value(row, "composition", ("Composition",))
+    composition = _composition_tags_text(deps.row_value(row, "composition", ("Composition",)))
     for value in (
         item,
         pose,
@@ -226,7 +233,7 @@ def hard_tags_result(request: SDXLPairTagRequest, deps: SDXLTagRouteDependencies
     hard_role = deps.clean(row.get("source_role_graph") or row.get("role_graph"))
     hard_clothing = deps.clean(root.get("hardcore_clothing_state"))
     expression = deps.clean(row.get("character_expression_text") or row.get("expression"))
-    composition = deps.clean(row.get("composition"))
+    composition = _composition_tags_text(deps.clean(row.get("composition")))
     for value in (
         hard_role,
         hard_item,
diff --git a/tools/prompt_route_simulation.py b/tools/prompt_route_simulation.py
index 022fcaa..9b1d984 100644
--- a/tools/prompt_route_simulation.py
+++ b/tools/prompt_route_simulation.py
@@ -569,6 +569,13 @@ def _composition_label_issues(name: str, prompts: dict[str, str]) -> list[str]:
     return issues
 
 
+def _sdxl_composition_tag_issues(name: str, sdxl_prompt: str) -> list[str]:
+    for tag in (part.strip() for part in sdxl_prompt.split(",")):
+        if re.search(r"\bcomposition$", tag, flags=re.IGNORECASE):
+            return [f"{name}.sdxl_prompt: composition_tag_label"]
+    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:
@@ -705,6 +712,7 @@ def _formatter_issues(
     issues.extend(_caption_expression_grammar_issues(name, caption_text))
     issues.extend(_krea_grammar_issues(name, krea_prompt))
     issues.extend(_composition_label_issues(name, prompts))
+    issues.extend(_sdxl_composition_tag_issues(name, sdxl_prompt))
 
     for label, value in (
         (f"{name}.krea_negative", krea.get("negative_prompt")),
diff --git a/tools/prompt_smoke.py b/tools/prompt_smoke.py
index 12f5383..1e813f6 100644
--- a/tools/prompt_smoke.py
+++ b/tools/prompt_smoke.py
@@ -4806,6 +4806,15 @@ def smoke_sdxl_tag_routes() -> None:
     _expect("27-year-old adult" in stale_character_route_tags, "SDXL route lost structured character age")
     _expect("warm olive skin" in stale_character_route_tags, "SDXL route lost structured character skin")
     _expect("stale" not in stale_character_route_tags, "SDXL route should not parse stale prompt character labels")
+    composition_label_row = _fixture_hardcore_row(
+        composition="coworking lounge frame with tall-window depth behind them composition",
+    )
+    composition_label_tags = sdxl_tag_routes.row_core_tags_result(
+        sdxl_tag_routes.SDXLRowTagRequest(composition_label_row, 1.29),
+        deps,
+    ).as_text()
+    _expect("tall-window depth behind them composition" not in composition_label_tags, "SDXL route kept raw composition label tag")
+    _expect("tall-window depth behind them" in composition_label_tags, "SDXL route lost composition detail while removing label")
     stale_prompt_row = _fixture_hardcore_row(
         prompt="stale raw prompt mentions fully nude naked pussy penis oral anal semen penetration",
         item="standing portrait setup",