fix(tagger): Moondream2 NL captions, tagger_mode names parity, remove Git LFS weights from cache

app.py

@@ -455,13 +455,39 @@ def on_artist_style_recommendations(style: str, lang: str):
 def _build_tagger_html(result: dict, lc: str) -> str:
     lines = []
     lines.append("<div class='whyx-tagger-panel'>")
-    lines.append(f"<div class='whyx-tagger-panel-title'>{t('tagger_ratings', lc)}</div>")
+
+    # ---- Natural-language caption (Moondream) --------------------------------
+    nl_caption = (result.get("nl_caption") or "").strip()
+    if nl_caption:
+        lines.append(
+            f"<div class='whyx-tagger-panel-title'>🗣️ {t('tagger_nl_caption', lc)}</div>"
+            f"<div class='whyx-tagger-nl-caption'>{html.escape(nl_caption)}</div>"
+        )
+
+    # ---- Pose summary ---------------------------------------------------------
+    pose_tags = result.get("pose_tags") or []
+    people = int(result.get("people_count") or 0)
+    if pose_tags or people:
+        pose_label = t("tagger_pose_single" if people == 1 else "tagger_pose_many", lc).format(n=people)
+        lines.append(f"<div class='whyx-tagger-panel-title' style='margin-top:10px;'>🧍 {t('tagger_pose_label', lc)}</div>")
+        if people:
+            lines.append(f"<div class='whyx-tagger-pose-people'>{html.escape(pose_label)}</div>")
+        if pose_tags:
+            lines.append(
+                "<div class='whyx-tagger-pose-tags'>"
+                + ", ".join(html.escape(p) for p in pose_tags)
+                + "</div>"
+            )
+
+    # ---- Rating distribution ---------------------------------------------------
+    lines.append(f"<div class='whyx-tagger-panel-title' style='margin-top:10px;'>{t('tagger_ratings', lc)}</div>")
     rating_colors = {"general": "#34D399", "sensitive": "#FBBF24", "questionable": "#F97316", "explicit": "#EF4444"}
     for name, score in result["ratings"].items():
         color = rating_colors.get(name, "var(--text-dim)")
         pct = int(score * 100)
         bar = "▰" * (pct // 10) + "▱" * (10 - pct // 10)
         lines.append(f"<div class='whyx-tagger-rating-row'><span class='whyx-tagger-rating-name' style='color:{color};'>{name}</span> <span class='whyx-tagger-rating-bar' style='color:{color};'>{bar}</span> <span class='whyx-tagger-rating-pct'>({score:.1%})</span></div>")
+
     if result["characters"]:
         lines.append(f"<div class='whyx-tagger-panel-title' style='margin-top:10px;'>{t('tagger_characters', lc)}</div>")
         lines.append("<div class='whyx-tagger-badges'>")
@@ -531,15 +557,30 @@ def on_tag_image(image, gen_threshold, char_threshold, fmt, lang, progress=gr.Pr
         html_out = _build_tagger_html(result, lc)
         general_names = list(result["general"].keys())[:60]
         char_names = list(result["characters"].keys())
-        general_choices = [(n.replace("_", " "), n) for n in general_names]
+
+        # Merge NL caption + pose tags into the tag chips so they can be
+        # picked up by the same apply/copy pipeline as booru tags.
+        extra_names: list[str] = []
+        if result.get("nl_caption"):
+            extra_names.append(result["nl_caption"].strip())
+        pose_tags = result.get("pose_tags") or []
+        extra_names.extend(pose_tags)
+
+        all_names = extra_names + general_names
+        general_choices = [(n, n) for n in all_names]
         char_choices = [(n.replace("_", " "), n) for n in char_names]
+
         raw_caption = result["caption"]
+        # Pose tags are folded into the caption list so the chips match the summary.
+        if pose_tags:
+            raw_caption = ", ".join([*pose_tags, raw_caption]) if raw_caption else ", ".join(pose_tags)
         esc_taglist = result["taglist"]
         tags_box = esc_taglist if fmt == "prompt" else raw_caption
+
         return (
             gr.update(value=html_out),
             gr.update(visible=True),
-            gr.update(choices=general_choices, value=general_names, visible=True),
+            gr.update(choices=general_choices, value=all_names, visible=True),
             gr.update(choices=char_choices, value=char_names, visible=True),
             gr.update(value=tags_box, visible=True),
             gr.update(visible=True),
@@ -1605,7 +1646,7 @@ with gr.Blocks(**_blocks_kw) as demo:
                                     (t("tagger_mode_eva", "en"), "wd:eva02"),
                                     (t("tagger_mode_swinv2", "en"), "wd:swinv2"),
                                     (t("tagger_mode_vit", "en"), "wd:vit"),
-                                    (t("tagger_mode_qwen_caption", "en"), "qwen"),
+                                    (t("tagger_mode_moondream_caption", "en"), "qwen"),
                                 ],
                                 value="ensemble",
                                 label="",

data/session_history.json.bak

@@ -1,4 +1,30 @@
 [
+  {
+    "id": "b5e660fd",
+    "timestamp": 1785823213.9892197,
+    "prompt": "1girl, blue hair",
+    "model": "anima",
+    "rating": "pg",
+    "num_variations": 2,
+    "creativity": "medium",
+    "weight_mode": "off",
+    "results": [
+      "masterpiece, best quality, score_7, safe, 1girl, BREAK, blue hair, feather wings, pastel colors, angel wings, surprised pose, cute pose, cinematic composition, circular composition, retro futuristic, rainbow, subsurface scattering, reflected moonlight, mysterious lighting, cloudy lighting, dreamy haze, glitch effect, beautiful detailed, beautiful detail",
+      "masterpiece, best quality, score_7, safe, 1girl, BREAK, blue hair, happy, gentle smile, from above, saturated matte, high contrast, vibrant, demon wings, profile view, detailed background, clock tower, mist, bioluminescent sea, underlit horror, moon light, silver aesthetic, blue tint, intricate details, beautiful detailed, beautiful detail",
+      "",
+      "",
+      "",
+      "",
+      "",
+      "",
+      "",
+      ""
+    ],
+    "neg_results": [],
+    "categories": [],
+    "seed": null,
+    "liked_indices": []
+  },
   {
     "id": "2dad208c",
     "timestamp": 1785821643.762218,
@@ -1272,31 +1298,5 @@
     "categories": [],
     "seed": null,
     "liked_indices": []
-  },
-  {
-    "id": "6f987ee2",
-    "timestamp": 1784831437.9085064,
-    "prompt": "1girl, blue hair",
-    "model": "anima",
-    "rating": "pg",
-    "num_variations": 2,
-    "creativity": "medium",
-    "weight_mode": "off",
-    "results": [
-      "masterpiece, best quality, score_9, safe, 1girl, BREAK, instagram filter, high contrast, blue hair, frilled dress, high heels, detailed background, wizard tower, phoenix, shield, sword, global illumination, gothic lighting, silver aesthetic, cosmic dust, floating, foreground framing, radial composition, first person view, looking at viewer, epic composition, surprised pose, from side, intricate details, beautiful detailed",
-      "masterpiece, best quality, score_9, safe, 1girl, BREAK, affectionate gaze, blue hair, crying, hoodie, detailed background, bullet train, motorcycle, skyscraper, dragon, global illumination, moonlight beam, holographic display, silver aesthetic, lightning aura, layered composition, surprised pose, looking away, incredibly absurdres, intricate details, magazine quality",
-      "",
-      "",
-      "",
-      "",
-      "",
-      "",
-      "",
-      ""
-    ],
-    "neg_results": [],
-    "categories": [],
-    "seed": null,
-    "liked_indices": []
   }
 ]

src/i18n.py

@@ -329,14 +329,13 @@ L10N = {
         "tagger_mode_eva": "WD14 EVA02",
         "tagger_mode_swinv2": "WD14 SwinV2",
         # --- tagger / ensemble ---
-        # --- tagger / ensemble ---
         "tagger_mode_vit": "WD14 ViT",
-        "tagger_mode_qwen": "Qwen3-VL (NL caption)",
-        "tagger_mode_qwen_desc": "Qwen3-VL (Natural-language caption)",
+        "tagger_mode_moondream_caption": "Moondream2 (detailed NL caption)",
         "tagger_pose_label": "Pose",
         "tagger_pose_single": "{n} person detected",
         "tagger_pose_many": "{n} people detected",
         "tagger_caption_label": "Natural-language caption",
+        "tagger_nl_caption": "🗣️ Natural-language caption",
         "tagger_chips_label": "Detected tags (click to toggle)",
         "tagger_chars_label": "Detected characters",
         "tagger_tags_label": "Tags to use",
@@ -700,8 +699,8 @@ L10N = {
         "tagger_mode_eva": "WD14 EVA02",
         "tagger_mode_swinv2": "WD14 SwinV2",
         "tagger_mode_vit": "WD14 ViT",
-        "tagger_mode_qwen": "Qwen3-VL (NL описание)",
-        "tagger_mode_qwen_desc": "Qwen3-VL (Подробное описание изображения)",
+        "tagger_mode_moondream_caption": "Moondream2 (подробное описание)",
+        "tagger_nl_caption": "🗣️ Описание сцены",
         "tagger_pose_label": "Поза",
         "tagger_pose_single": "Обнаружен {n} человек",
         "tagger_pose_many": "Обнаружено {n} человек",