fix video pipeline: timecode sync, face framing in prompts, no grouping when timecodes present

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

story-gen/scripts/generate.py

@@ -117,6 +117,15 @@ Rules for scene structure fields:
 - outcome: what this scene resolves or leads to (null if opening scene)
 - story_value: narrative purpose — one of: hook|problem|feature|climax|resolution|cta|transition
 
+Rules for timecode and duration:
+- timecode: MUST be the cumulative start time of this scene (HH:MM:SS). Scene 1 starts at 00:00:00. Each next scene starts at previous timecode + previous duration_sec.
+- duration_sec: MUST exactly match the length of the voiceover text. Count ~2.5 words per second for Russian, ~3 words per second for English.
+- voiceover and duration_sec MUST be in sync: if voiceover has 15 words in Russian → duration_sec = 6.
+
+Rules for action (visual description):
+- If the video involves people or faces: ALWAYS describe a tight face/portrait shot unless the scene specifically needs wide frame. Example: "Close-up of man's face, intense expression, dramatic lighting"
+- Describe the main subject clearly — camera framing (close-up, medium shot, wide), subject action, lighting
+
 Response format:
 {
   "title": "video title",
@@ -137,7 +146,7 @@ Response format:
       "outcome": "what this resolves or leads to, or null",
       "story_value": "hook|problem|feature|climax|resolution|cta|transition",
       "voiceover": "exact words spoken by narrator in target language",
-      "action": "detailed English description of what is on screen: camera, subject, movement, lighting"
+      "action": "detailed English description of what is on screen: camera framing, subject, movement, lighting"
     }
   ],
   "music_mood": "upbeat|calm|dramatic|funny|inspirational",

voice/tts_generate.py

@@ -88,37 +88,30 @@ def parse_json_transcription(json_path):
         combined_text = ". ".join(texts).replace("\n", " ").replace("\\n", " ")
         print(f"  Объединённый текст: {combined_text[:100]}...")
 
-        # Определяем время
-        if "selected_timestamp" in data.get("source", {}):
-            # Используем таймкод из source
-            time_str = data["source"]["selected_timestamp"]
-            print(f"  Время из source.selected_timestamp: {time_str}")
-            parts = time_str.split(':')
+        # Определяем время — приоритет: timecode (story-gen video mode) > start_time > source > i*duration
+        def parse_timecode(tc: str) -> float:
+            parts = tc.strip().split(':')
             if len(parts) == 3:
-                hours, minutes, seconds = parts
-                total_seconds = int(hours) * 3600 + int(minutes) * 60 + float(seconds)
+                return int(parts[0]) * 3600 + int(parts[1]) * 60 + float(parts[2])
             elif len(parts) == 2:
-                minutes, seconds = parts
-                total_seconds = int(minutes) * 60 + float(seconds)
-            else:
-                total_seconds = float(parts[0])
-        elif "start_time" in scene:
-            start_time_str = scene["start_time"]
-            print(f"  Время из scene.start_time: {start_time_str}")
-            # Парсим время в секундах
-            parts = start_time_str.split(':')
-            if len(parts) == 3:
-                hours, minutes, seconds = parts
-                total_seconds = int(hours) * 3600 + int(minutes) * 60 + float(seconds)
-            elif len(parts) == 2:
-                minutes, seconds = parts
-                total_seconds = int(minutes) * 60 + float(seconds)
-            else:
-                total_seconds = float(parts[0])
+                return int(parts[0]) * 60 + float(parts[1])
+            return float(parts[0])
+
+        if "timecode" in scene and scene["timecode"]:
+            total_seconds = parse_timecode(scene["timecode"])
+            print(f"  Время из scene.timecode: {scene['timecode']} → {total_seconds}с")
+        elif "start_time" in scene and scene["start_time"]:
+            total_seconds = parse_timecode(scene["start_time"])
+            print(f"  Время из scene.start_time: {scene['start_time']} → {total_seconds}с")
+        elif "selected_timestamp" in data.get("source", {}):
+            total_seconds = parse_timecode(data["source"]["selected_timestamp"])
+            print(f"  Время из source.selected_timestamp: {total_seconds}с")
         else:
-            # Если нет времени, используем индекс сцены * 5 секунд
-            total_seconds = i * 5
-            print(f"  Время не найдено, используем индекс: {total_seconds}с")
+            # Fallback: накапливаем по duration_sec предыдущих сцен
+            total_seconds = sum(
+                s.get("duration_sec", 5) for s in scenes[:i]
+            )
+            print(f"  Время вычислено по duration_sec: {total_seconds}с")
 
         entries.append((total_seconds, combined_text, scene_id))
         print(f"  {GREEN}✓ Добавлена фраза для озвучки{RESET}")
@@ -201,8 +194,19 @@ async def main_async(json_path, output_dir, voice, max_chars):
     for ts, text, scene_id in entries:
         print(f"  - [{ts:.2f}с] Сцена {scene_id}: {text[:50]}...")
 
-    print(f"\n{YELLOW}Группировка фраз (макс. {max_chars} символов)...{RESET}")
-    grouped_entries = group_sentences(entries, max_chars=max_chars)
+    # Если сцены имеют точные таймкоды — не группируем, каждая сцена отдельный сегмент
+    try:
+        _raw = json.loads(Path(json_path).read_text(encoding='utf-8'))
+        has_timecodes = any(s.get("timecode") for s in _raw.get("scenes", []))
+    except Exception:
+        has_timecodes = False
+
+    if has_timecodes:
+        print(f"{GREEN}Режим точных таймкодов — каждая сцена отдельный сегмент.{RESET}")
+        grouped_entries = [(ts, text, [sid]) for ts, text, sid in entries]
+    else:
+        print(f"\n{YELLOW}Группировка фраз (макс. {max_chars} символов)...{RESET}")
+        grouped_entries = group_sentences(entries, max_chars=max_chars)
     print(f"{GREEN}Сформировано {len(grouped_entries)} фрагментов для озвучки.{RESET}")
 
     manifest_path = output_path / "segments.txt"