Split SDXL sentence-boundary tags

This commit is contained in:
2026-06-27 22:23:49 +02:00
parent 842d3580f5
commit 030a1255e1
3 changed files with 16 additions and 0 deletions
+1
View File
@@ -63,6 +63,7 @@ def split_tag_text(text: Any) -> list[str]:
text,
flags=re.IGNORECASE,
)
text = re.sub(r"(?<=[A-Za-z0-9)])\.\s+(?=[A-Za-z])", ", ", text)
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)
+7
View File
@@ -594,6 +594,12 @@ def _sdxl_character_pair_fragment_issues(name: str, sdxl_prompt: str) -> list[st
return []
def _sdxl_sentence_boundary_tag_issues(name: str, sdxl_prompt: str) -> list[str]:
if re.search(r"\b[a-z][^,.]{0,80}\.\s+(?:woman|man|the|keep|use)\b", sdxl_prompt, flags=re.IGNORECASE):
return [f"{name}.sdxl_prompt: sentence_boundary_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:
@@ -734,6 +740,7 @@ def _formatter_issues(
issues.extend(_sdxl_expression_label_issues(name, sdxl_prompt))
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))
for label, value in (
(f"{name}.krea_negative", krea.get("negative_prompt")),
+8
View File
@@ -4683,6 +4683,14 @@ def smoke_sdxl_tag_policy() -> None:
subject_pair_tags = sdxl_tag_policy.split_tag_text("Woman A, Man A are mid-transition with hands on hips")
_expect("woman and man are mid-transition" in subject_pair_tags, "SDXL tag splitter broke paired character clause")
_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("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(
"thighs. man watches close" not in sentence_boundary_tags_lower,
"SDXL tag splitter kept sentence-boundary tag fragment",
)
_expect(
sdxl_formatter._camera_tags(row) == sdxl_tag_policy.camera_tags(row),
"SDXL formatter camera helper should delegate to sdxl_tag_policy",