fix: Use LLM to generate unique scene prompts for video extensions

app/services/grok/services/video.py

@@ -2,6 +2,7 @@
 Grok video generation service.
 """
 
+import aiohttp
 import asyncio
 import math
 import re
@@ -40,6 +41,56 @@
 _POST_ID_URL_PATTERN = r"/generated/([0-9a-fA-F-]{32,36})/"
 
 
+async def _generate_scene_prompts_llm(original_prompt: str, num_scenes: int) -> List[str]:
+    """Use local Grok API to generate unique scene prompts for each video round."""
+
+    num_scenes = min(num_scenes, 6)
+
+    system_msg = f"""Break this video concept into {num_scenes} sequential scenes that flow naturally.
+
+Original concept: "{original_prompt}"
+
+CRITICAL RULES:
+- Each scene MUST be DIFFERENT (no repetition)
+- Scenes progress naturally like a story
+- Smooth transitions between scenes
+- Each scene continues from the previous one
+- Output ONLY a JSON array: ["scene 1 description", "scene 2 description", ...]
+
+Generate exactly {num_scenes} unique scenes:"""
+
+    try:
+        async with aiohttp.ClientSession() as session:
+            async with session.post(
+                "http://localhost:8000/v1/chat/completions",
+                headers={"Content-Type": "application/json"},
+                json={
+                    "model": "grok-4.1-fast",
+                    "messages": [{"role": "user", "content": system_msg}],
+                    "temperature": 0.8,
+                    "max_tokens": 2000
+                },
+                timeout=aiohttp.ClientTimeout(total=30)
+            ) as resp:
+                if resp.status == 200:
+                    data = await resp.json()
+                    content = data.get("choices", [{}])[0].get("message", {}).get("content", "")
+                    match = re.search(r'\[[\s\S]*?\]', content)
+                    if match:
+                        scenes = orjson.loads(match.group(0))
+                        if isinstance(scenes, list) and len(scenes) >= num_scenes:
+                            logger.info(f"Generated {len(scenes)} unique scene prompts via LLM")
+                            return [str(s) for s in scenes[:num_scenes]]
+                else:
+                    logger.warning(f"LLM API error {resp.status}")
+    except Exception as e:
+        logger.warning(f"LLM scene generation failed: {e}")
+
+    # Fallback to original prompt with scene numbers
+    logger.info(f"Using fallback scene prompts")
+    return [f"{original_prompt} (scene {i+1}/{num_scenes})" for i in range(num_scenes)]
+
+
 @dataclass(frozen=True)
 class VideoRoundPlan:
     round_index: int
@@ -811,6 +862,10 @@ async def completions(
         round_plan = _build_round_plan(target_length, is_super=is_super_pool)
         total_rounds = len(round_plan)
 
+        # Generate unique scene prompts for each round using LLM
+        scene_prompts = await _generate_scene_prompts_llm(prompt, total_rounds)
+        logger.info(f"Using {len(scene_prompts)} scene prompts for {total_rounds} rounds")
+
         service = VideoService()
         message = _build_message(prompt, preset)
 
@@ -849,18 +904,22 @@ async def _run_round_collect(
             original_id: Optional[str],
             source: str,
         ) -> VideoRoundResult:
+            # Use unique scene prompt for this round
+            round_prompt = scene_prompts[plan.round_index - 1] if plan.round_index <= len(scene_prompts) else prompt
+            round_message = _build_message(round_prompt, preset)
+
             config_override = _build_round_config(
                 plan,
                 seed_post_id=seed_id,
                 last_post_id=last_id,
                 original_post_id=original_id,
-                prompt=prompt,
+                prompt=round_prompt,
                 aspect_ratio=aspect_ratio,
                 resolution_name=generation_resolution,
             )
             response = await _request_round_stream(
                 token=token,
-                message=message,
+                message=round_message,
                 model_config_override=config_override,
             )
             return await _collect_round_result(response, model=model, source=source)
@@ -874,18 +933,22 @@ async def _stream_chain() -> AsyncGenerator[str, None]:
 
             try:
                 for plan in round_plan:
+                    # Use unique scene prompt for this round
+                    round_prompt = scene_prompts[plan.round_index - 1] if plan.round_index <= len(scene_prompts) else prompt
+                    round_message = _build_message(round_prompt, preset)
+
                     config_override = _build_round_config(
                         plan,
                         seed_post_id=seed_id,
                         last_post_id=last_id,
                         original_post_id=original_id,
-                        prompt=prompt,
+                        prompt=round_prompt,
                         aspect_ratio=aspect_ratio,
                         resolution_name=generation_resolution,
                     )
                     response = await _request_round_stream(
                         token=token,
-                        message=message,
+                        message=round_message,
                         model_config_override=config_override,
                     )
 

app/services/grok/services/video_extend.py

@@ -2,10 +2,13 @@
 Direct video extension service (app-chat based).
 """
 
+import aiohttp
 import re
 import time
 import uuid
-from typing import Any, Dict, Optional
+from typing import Any, Dict, List, Optional
+
+import orjson
 
 from app.core.exceptions import AppException, ErrorType, UpstreamException, ValidationException
 from app.core.logger import logger
@@ -18,6 +21,51 @@
 
 VIDEO_MODEL_ID = "grok-imagine-1.0-video"
 
+
+async def _generate_scene_prompt_for_extend(original_prompt: str, current_scene: int, total_scenes: int) -> str:
+    """Use local Grok API to generate unique scene prompt for video extension."""
+
+    system_msg = f"""Continue this video concept with scene {current_scene} of {total_scenes}.
+
+Original concept: "{original_prompt}"
+
+CRITICAL RULES:
+- Scene MUST continue from previous scene
+- Natural progression, NO repetition
+- Different angle/action from previous
+- Output ONLY the scene description (no JSON, no quotes)
+
+Generate scene {current_scene}:"""
+
+    try:
+        async with aiohttp.ClientSession() as session:
+            async with session.post(
+                "http://localhost:8000/v1/chat/completions",
+                headers={"Content-Type": "application/json"},
+                json={
+                    "model": "grok-4.1-fast",
+                    "messages": [{"role": "user", "content": system_msg}],
+                    "temperature": 0.8,
+                    "max_tokens": 300
+                },
+                timeout=aiohttp.ClientTimeout(total=20)
+            ) as resp:
+                if resp.status == 200:
+                    data = await resp.json()
+                    content = data.get("choices", [{}])[0].get("message", {}).get("content", "").strip()
+                    if content:
+                        logger.info(f"Generated extend scene {current_scene}/{total_scenes} via LLM")
+                        return content
+                else:
+                    logger.warning(f"LLM API error {resp.status}")
+    except Exception as e:
+        logger.warning(f"LLM extend scene generation failed: {e}")
+
+    # Fallback
+    logger.info(f"Using fallback extend prompt")
+    return f"{original_prompt} (continuation {current_scene}/{total_scenes})"
+
+
 _RATIO_MAP = {
     "1280x720": "16:9",
     "720x1280": "9:16",
@@ -120,6 +168,14 @@ async def extend(
             )
         resolution_name = _normalize_resolution(resolution)
 
+        # LLM ile extend için sahne promptu oluştur
+        current_scene = int(start_time / 6) + 2  # +2 çünkü base video 1. sahne
+        total_scenes = 6  # Max 6 sahne (30 saniye)
+
+        logger.info(f"🎬 EXTEND: start_time={start_time}s → scene {current_scene}/{total_scenes}")
+        extend_prompt = await _generate_scene_prompt_for_extend(prompt, current_scene, total_scenes)
+        logger.info(f"🎥 EXTEND scene {current_scene}: {extend_prompt[:150]}...")
+
         token_mgr = await get_token_manager()
         await token_mgr.reload_if_stale()
 
@@ -147,7 +203,7 @@ async def extend(
                     "videoExtensionStartTime": float(start_time),
                     "extendPostId": reference_id,
                     "stitchWithExtendPostId": True,
-                    "originalPrompt": prompt,
+                    "originalPrompt": extend_prompt,
                     "originalPostId": reference_id,
                     "originalRefType": "ORIGINAL_REF_TYPE_VIDEO_EXTENSION",
                     "mode": "custom",
@@ -165,7 +221,7 @@ async def extend(
         response = await AppChatReverse.request(
             session,
             token,
-            message=f"{prompt} --mode=custom",
+            message=f"{extend_prompt} --mode=custom",
             model="grok-3",
             tool_overrides={"videoGen": True},
             model_config_override=model_config_override,
@@ -200,7 +256,7 @@ async def extend(
             "created_at": now,
             "completed_at": now,
             "status": "completed",
-            "prompt": prompt,
+            "prompt": extend_prompt,
             "reference_id": reference_id,
             "start_time": float(start_time),
             "ratio": aspect_ratio,