feat: ML-Danbooru primary runs only route to loli

Co-authored-by: Cursor <cursoragent@cursor.com>

backend/app/api.py

@@ -50,6 +50,7 @@ from .services import (
     resolve_settings,
     sanitize_folder_name,
     categories_exclude_dirs,
+    destination_folders_for_tagger,
     scan_images,
 )
 from .taxonomy import (
@@ -1895,6 +1896,9 @@ def start_run(payload: StartRunRequest) -> StartRunResponse:
         selected_folders = payload.selected_folders
         if not selected_folders:
             selected_folders = list(current.selected_tags)
+        selected_folders = destination_folders_for_tagger(
+            selected_folders, current.tagger_model
+        )
 
     try:
         known_tags = load_known_tags(TAGS_CSV)
@@ -1962,6 +1966,8 @@ def start_run(payload: StartRunRequest) -> StartRunResponse:
         if real_life_filter
         else ""
     )
+    if not real_life_filter and current.tagger_model == "ml_danbooru":
+        mode_note += " ML-Danbooru mode: loli destination only."
     return StartRunResponse(
         run_id=run_id,
         status="pending",
@@ -2038,6 +2044,9 @@ def reclassify_run(run_id: int, payload: ReclassifyRequest) -> ReclassifyRespons
             status_code=400,
             detail="No selected tags in settings; save destination tags before reclassifying.",
         )
+    selected_folders = destination_folders_for_tagger(
+        selected_folders, payload.tagger_model
+    )
     try:
         known_tags = load_known_tags(TAGS_CSV)
         mappings = discover_tag_folders(categories_root, known_tags, selected_folders)

backend/app/services.py

@@ -665,3 +665,21 @@ def resolve_settings(
             ),
         }
     )
+
+
+def destination_folders_for_tagger(
+    selected_folders: list[str],
+    tagger_model: str,
+    *,
+    real_life_filter: bool = False,
+) -> list[str]:
+    """ML-Danbooru primary runs only compete for the loli destination.
+
+    Benchmarks showed strong loli recall for ML, but weak/noisy behavior on the
+    broader preferred-folder set — so full multi-folder routing stays on WD.
+    """
+    if real_life_filter:
+        return list(selected_folders)
+    if str(tagger_model or "") == "ml_danbooru":
+        return ["loli"]
+    return list(selected_folders)

backend/tests/test_ml_loli_only.py

@@ -0,0 +1,14 @@
+"""ML-Danbooru primary runs are loli-only (strong recall on that folder)."""
+
+from __future__ import annotations
+
+from app.services import destination_folders_for_tagger
+
+
+def test_ml_danbooru_forces_loli_only_destinations() -> None:
+    selected = ["Voyeur", "fellatio", "loli", "monster_girl"]
+    assert destination_folders_for_tagger(selected, "ml_danbooru") == ["loli"]
+    assert destination_folders_for_tagger(selected, "wd_swinv2_v3") == selected
+    assert destination_folders_for_tagger(
+        ["real_life"], "ml_danbooru", real_life_filter=True
+    ) == ["real_life"]

frontend/src/App.jsx

@@ -23,7 +23,7 @@ const TAGGER_MODELS = [
   {
     id: "ml_danbooru",
     label: "ML-Danbooru",
-    help: "Original ONNX tagger used by this app. Fast baseline.",
+    help: "Loli-only classifier — strong recall on that folder (not used for full multi-folder runs).",
   },
   {
     id: "wd_swinv2_v3",