From 718da9a68dc7c32826017f4abafafa8b75d39d9c Mon Sep 17 00:00:00 2001
From: Ethanfel <ethan.fel@ts-pc.fr>
Date: Sat, 27 Jun 2026 22:27:41 +0200
Subject: [PATCH] Clean SDXL imperative hand tags

---
 sdxl_tag_policy.py               | 7 ++++++-
 tools/prompt_route_simulation.py | 7 +++++++
 tools/prompt_smoke.py            | 2 ++
 3 files changed, 15 insertions(+), 1 deletion(-)

diff --git a/sdxl_tag_policy.py b/sdxl_tag_policy.py
index 1db7041..038adf9 100644
--- a/sdxl_tag_policy.py
+++ b/sdxl_tag_policy.py
@@ -67,7 +67,12 @@ def split_tag_text(text: Any) -> list[str]:
     text = re.sub(r"(?<!-)\b(?:and|with)\b(?!-)", ",", text, flags=re.IGNORECASE)
     text = re.sub(r"\b(woman|man),\s+(woman|man)\s+are\b", r"\1 and \2 are", text, flags=re.IGNORECASE)
     parts = re.split(r"\s*[,;]\s*", text)
-    return [clean(part).strip(" .") for part in parts if clean(part).strip(" .")]
+    tags = []
+    for part in parts:
+        part = re.sub(r"^keep\s+", "", clean(part).strip(" ."), flags=re.IGNORECASE)
+        if part:
+            tags.append(part)
+    return tags
 
 
 def tag_key(tag: str) -> str:
diff --git a/tools/prompt_route_simulation.py b/tools/prompt_route_simulation.py
index b590974..6445191 100644
--- a/tools/prompt_route_simulation.py
+++ b/tools/prompt_route_simulation.py
@@ -600,6 +600,12 @@ def _sdxl_sentence_boundary_tag_issues(name: str, sdxl_prompt: str) -> list[str]
     return []
 
 
+def _sdxl_imperative_tag_issues(name: str, sdxl_prompt: str) -> list[str]:
+    if re.search(r"\bkeep\s+hands\b", sdxl_prompt, flags=re.IGNORECASE):
+        return [f"{name}.sdxl_prompt: imperative_keep_hands_tag"]
+    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:
@@ -741,6 +747,7 @@ def _formatter_issues(
     issues.extend(_sdxl_hyphen_fragment_issues(name, sdxl_prompt))
     issues.extend(_sdxl_character_pair_fragment_issues(name, sdxl_prompt))
     issues.extend(_sdxl_sentence_boundary_tag_issues(name, sdxl_prompt))
+    issues.extend(_sdxl_imperative_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 3eb8a64..0a33727 100644
--- a/tools/prompt_smoke.py
+++ b/tools/prompt_smoke.py
@@ -4685,6 +4685,8 @@ def smoke_sdxl_tag_policy() -> None:
     _expect("woman" not in subject_pair_tags and "man are mid-transition" not in subject_pair_tags, "SDXL tag splitter emitted broken paired character fragments")
     sentence_boundary_tags = sdxl_tag_policy.split_tag_text("keep hands on hips, breasts, thighs. Man watches close")
     sentence_boundary_tags_lower = [tag.lower() for tag in sentence_boundary_tags]
+    _expect("hands on hips" in sentence_boundary_tags_lower, "SDXL tag splitter did not clean leading keep imperative")
+    _expect("keep hands on hips" not in sentence_boundary_tags_lower, "SDXL tag splitter kept leading keep imperative")
     _expect("thighs" in sentence_boundary_tags, "SDXL tag splitter lost pre-period tag")
     _expect("man watches close" in sentence_boundary_tags_lower, "SDXL tag splitter did not split sentence-boundary tag")
     _expect(