Update core/prompt_engineering.py

core/prompt_engineering.py

@@ -1,38 +1,37 @@
 # core/prompt_engineering.py
 import json
 
-# create_story_breakdown_prompt can remain mostly the same, but ensure 'key_action' is descriptive.
-# If key_action is very short, text overlays on video might not be useful.
-# Consider making 'key_action' slightly more verbose or adding a 'video_overlay_text' field.
-# For now, we'll use key_action.
-
 def create_story_breakdown_prompt(user_idea, genre="sci-fi", mood="suspenseful", num_scenes=3):
+    """
+    Generates a prompt for Gemini to break down a story idea into scenes.
+    Emphasizes concise, descriptive key_actions suitable for video overlays.
+    """
     return f"""
     You are an expert screenwriter and visual storyteller.
     Based on the user's idea: "{user_idea}"
     And considering the genre: "{genre}" and mood: "{mood}"
 
     Generate a {num_scenes}-scene story breakdown. For each scene, provide:
-    1. scene_number (int)
-    2. setting_description (str): Vivid description of the location and atmosphere (approx 20-40 words).
-    3. characters_involved (list of str): Names of characters in the scene.
-    4. key_action (str): The main event or action happening in one concise sentence (approx 10-20 words). This will be used for video overlays.
-    5. dialogue_snippet (str): A brief, impactful line of dialogue if applicable.
-    6. visual_style_suggestion (str): e.g., "Dark and gritty, high contrast, Blade Runner-esque neon"
-    7. camera_angle_suggestion (str): e.g., "Low-angle shot to emphasize power"
-    8. emotional_beat (str): The core emotion or turning point in the scene.
+    1. scene_number (int): Sequential number of the scene.
+    2. emotional_beat (str): A short title or phrase capturing the core emotion/theme of this scene (e.g., "Desperate Escape," "Betrayal Revealed," "A Glimmer of Hope").
+    3. setting_description (str): Vivid description of the location, atmosphere, and key visual elements (approx 30-50 words).
+    4. characters_involved (list of str): Names of characters present and active in the scene.
+    5. key_action (str): The single most important event or character action happening in this specific visual moment, described concisely for a potential video overlay (max 15-20 words).
+    6. dialogue_snippet (str): A brief, impactful line of dialogue spoken in this scene, if any.
+    7. visual_style_suggestion (str): Keywords for the overall visual style of this scene (e.g., "Dark and gritty, high contrast, Blade Runner-esque neon reflections"). This can be influenced by the overall story mood.
+    8. camera_angle_suggestion (str): A specific camera shot type or angle suggestion (e.g., "Low-angle shot emphasizing power," "Dutch angle for unease," "Extreme close-up on eyes").
 
     Output ONLY the JSON object for the list of scenes.
     Example for one scene:
     {{
       "scene_number": 1,
-      "setting_description": "A dimly lit, cluttered spaceship cockpit. Warning lights flash intermittently. Steam vents from a broken pipe.",
-      "characters_involved": ["Captain Eva Rostova"],
-      "key_action": "Eva frantically works at a console, trying to divert a catastrophic system failure.",
-      "dialogue_snippet": "Eva: 'Come on, come on... don't do this to me now!'",
-      "visual_style_suggestion": "Claustrophobic, practical lighting, lens flares, metallic sheens.",
-      "camera_angle_suggestion": "Close-up on Eva's determined face, sweat beading on her forehead.",
-      "emotional_beat": "Desperation and intense focus."
+      "emotional_beat": "Tense Standoff",
+      "setting_description": "A rain-slicked, neon-drenched alleyway in Neo-Kyoto. Broken holographic advertisements flicker erratically, casting distorted shadows. The air hangs heavy with the smell of ozone and despair.",
+      "characters_involved": ["Detective Kaito Tanaka", "Informant (shadowy figure)"],
+      "key_action": "Kaito cautiously approaches a nervous informant huddled in the shadows.",
+      "dialogue_snippet": "Informant: 'They know you're looking... You're not safe.'",
+      "visual_style_suggestion": "Neo-noir, cyberpunk, high contrast, deep shadows, vibrant neon reflections in puddles, film grain.",
+      "camera_angle_suggestion": "Medium shot from behind Kaito, focusing on the informant, creating suspense."
     }}
 
     Provide the full JSON structure for {num_scenes} scenes in a list:
@@ -43,91 +42,140 @@ def create_story_breakdown_prompt(user_idea, genre="sci-fi", mood="suspenseful",
     ]
     """
 
-def create_image_prompt_from_scene_data(scene_data, character_definitions=None, style_reference_desc=None):
+def create_image_prompt_from_scene_data(scene_data, character_definitions=None, global_style_reference=""):
     """
-    Generates an image prompt from structured scene data, injecting character details.
-    scene_data: dictionary for a single scene.
-    character_definitions: dict {char_name_lower: description} for consistency.
-    style_reference_desc: textual description of a desired style.
+    Generates a detailed image prompt for DALL-E 3 based on structured scene data,
+    injecting character descriptions and global style preferences.
+
+    scene_data (dict): Contains details for a single scene.
+    character_definitions (dict): {'character_name_lower': 'description', ...}
+    global_style_reference (str): User-defined global style keywords.
     """
-    setting_desc = scene_data.get('setting_description', '')
-    key_action_desc = scene_data.get('key_action', '')
-    visual_style = scene_data.get('visual_style_suggestion', 'cinematic')
-    camera_angle = scene_data.get('camera_angle_suggestion', '')
-    emotional_beat = scene_data.get('emotional_beat', '')
+    scene_num = scene_data.get('scene_number', 'N/A')
+    emotional_beat_title = scene_data.get('emotional_beat', 'A cinematic scene') # Used as part of the core request
+    setting_desc = scene_data.get('setting_description', 'A visually interesting setting.')
+    key_action_desc = scene_data.get('key_action', 'A significant moment unfolds.')
     
-    characters_str_parts = []
-    if character_definitions and scene_data.get('characters_involved'):
-        for char_name in scene_data.get('characters_involved', []):
-            # Match character name case-insensitively, but use original name from scene_data
-            defined_desc = character_definitions.get(char_name.lower().strip())
-            if defined_desc:
-                characters_str_parts.append(f"{char_name.strip()} ({defined_desc})")
+    # --- Character Injection ---
+    characters_involved_in_scene = scene_data.get('characters_involved', [])
+    character_prompt_segments = []
+    if characters_involved_in_scene:
+        for char_name_from_scene in characters_involved_in_scene:
+            char_name_clean = char_name_from_scene.strip()
+            char_lookup_key = char_name_clean.lower()
+            if character_definitions and char_lookup_key in character_definitions:
+                char_visual_desc = character_definitions[char_lookup_key]
+                character_prompt_segments.append(f"{char_name_clean} (described as: {char_visual_desc})")
             else:
-                characters_str_parts.append(char_name.strip()) # Character mentioned but not defined
-    elif scene_data.get('characters_involved'): # Characters mentioned but no definitions provided
-        characters_str_parts = [name.strip() for name in scene_data.get('characters_involved', [])]
-
-    characters_involved_str = ""
-    if characters_str_parts:
-        if len(characters_str_parts) == 1:
-            characters_involved_str = f" The scene features {characters_str_parts[0]}."
+                character_prompt_segments.append(char_name_clean) # Character is present but no specific visual definition provided
+    
+    characters_narrative = ""
+    if character_prompt_segments:
+        if len(character_prompt_segments) == 1:
+            characters_narrative = f"The primary focus is on {character_prompt_segments[0]}."
         else:
-            characters_involved_str = f" The scene features {', '.join(characters_str_parts[:-1])} and {characters_str_parts[-1]}."
+            characters_narrative = f"The scene prominently features {', '.join(character_prompt_segments[:-1])} and {character_prompt_segments[-1]}."
+    # --- End Character Injection ---
+
+    # --- Style Aggregation ---
+    scene_specific_style = scene_data.get('visual_style_suggestion', 'cinematic, photorealistic')
+    final_style_directive = scene_specific_style
+    if global_style_reference: # User's global style preference
+        final_style_directive += f", {global_style_reference}"
+    # --- End Style Aggregation ---
+
+    camera_instr = scene_data.get('camera_angle_suggestion', 'eye-level medium shot')
 
-    base_desc = f"Depict: Scene {scene_data.get('scene_number', '')}. {key_action_desc} {characters_involved_str} Setting: {setting_desc}."
+    # Constructing the DALL-E 3 prompt with more narrative flow
+    # DALL-E 3 often works best if you tell it what kind of image you want first, then the details.
+    prompt = (
+        f"Create an ultra-detailed, photorealistic, and highly cinematic digital painting or concept art image. "
+        f"The image should depict: '{emotional_beat_title}'. "
+        f"Setting: {setting_desc}. "
+        f"{characters_narrative} " # This will be empty if no characters, or list them with descriptions
+        f"They are engaged in the following key action: {key_action_desc}. "
+        f"Visual Style and Atmosphere: {final_style_directive}. "
+        f"Camera Composition: {camera_instr}. "
+        f"Emphasize: Dramatic lighting (consider {scene_data.get('mood','cinematic')} mood), rich textures, depth of field, and strong atmospheric effects like mist, rain, or dust if appropriate to the setting. "
+        f"The overall image must feel like a high-quality still from a major motion picture or a AAA video game. "
+        f"Pay close attention to character details if provided, ensuring they are distinct and match their descriptions."
+    )
     
-    style_mod = ""
-    if style_reference_desc:
-        style_mod = f" Artistic style inspired by: {style_reference_desc}."
-
-    # Constructing a more robust prompt for DALL-E 3
-    # DALL-E 3 benefits from descriptive, story-like prompts.
-    full_prompt = f"""
-    Generate a highly detailed, photorealistic and cinematic image.
-    Image Description: {base_desc}
-    Visual Style: {visual_style}. {style_mod}
-    Camera Perspective: {camera_angle}.
-    Emotional Tone: {emotional_beat}.
-    Key elements to emphasize: Cinematic composition, dramatic lighting, depth of field, rich textures, and atmospheric effects.
-    Output a visually stunning image suitable for a film storyboard.
-    """
-    # Removed "Output only the prompt string" as this function *is* the prompt string builder.
-    # DALL-E 3 does not need "Suitable for DALL-E 3 or Midjourney".
-    return " ".join(full_prompt.split()) # Cleans up extra whitespace
+    return " ".join(prompt.split()) # Normalize whitespace
 
 def create_scene_regeneration_prompt(original_scene_data, user_feedback, full_story_context=None):
-    # ... (remains the same as your last working version) ...
-    context_str = f"Original scene details:\n{json.dumps(original_scene_data, indent=2)}\n\n"
+    """
+    Prompt for Gemini to regenerate scene script details based on user feedback.
+    """
+    context_str = f"Original scene (Scene Number {original_scene_data.get('scene_number')}):\n{json.dumps(original_scene_data, indent=2)}\n\n"
     if full_story_context:
-        context_str += f"Full story context (this scene is number {original_scene_data.get('scene_number')}):\n{json.dumps(full_story_context, indent=2)}\n\n"
+        context_str += f"Full story context for reference:\n{json.dumps(full_story_context, indent=2)}\n\n"
+
     return f"""
-    You are an expert script doctor.
+    You are an expert script doctor and editor.
     {context_str}
-    The user wants to modify this scene based on the following feedback: "{user_feedback}"
+    The user wants to modify this specific scene based on the following feedback: "{user_feedback}"
+
     Please regenerate ONLY the JSON object for this single scene, incorporating the feedback.
-    Maintain the same JSON structure as the original scene (scene_number, setting_description, characters_involved, key_action, dialogue_snippet, visual_style_suggestion, camera_angle_suggestion, emotional_beat).
-    Ensure the scene_number remains the same. The 'key_action' should be a concise descriptive sentence (10-20 words).
+    Maintain the exact same JSON structure as the original: (scene_number, emotional_beat, setting_description, characters_involved, key_action, dialogue_snippet, visual_style_suggestion, camera_angle_suggestion).
+    The 'scene_number' must remain unchanged.
+    The 'key_action' should be a concise descriptive sentence (max 15-20 words) suitable for a brief video overlay.
+    If feedback pertains to characters, setting, action, dialogue, style, or camera, update those fields accordingly.
+    Ensure the regenerated scene remains coherent with the overall story context if provided.
+    Focus on making the changes impactful and clear.
     """
 
-def create_visual_regeneration_prompt(original_image_prompt, user_feedback_on_visuals, scene_data):
-    # ... (remains the same, this prompt is for Gemini to *rewrite* an image prompt) ...
-    # This prompt should also be updated to be DALL-E 3 friendly if it's rewriting.
-    return f"""
-    The previous detailed image generation prompt for a scene was: "{original_image_prompt}"
-    The scene details are:
-    Setting: {scene_data.get('setting_description')}
-    Action: {scene_data.get('key_action')}
-    Characters: {', '.join(scene_data.get('characters_involved',[]))}
-    Mood/Emotion: {scene_data.get('emotional_beat')}
-    Current Visual Style: {scene_data.get('visual_style_suggestion')}
-    Current Camera: {scene_data.get('camera_angle_suggestion')}
-
-    The user provided this feedback on the visuals: "{user_feedback_on_visuals}"
-
-    Generate a new, refined, highly detailed, photorealistic and cinematic image generation prompt based on this feedback.
-    The new prompt should aim to correct or enhance the visuals as per the user's request, while maintaining the core scene elements.
-    Ensure the prompt is descriptive and suitable for generating a stunning image for a film storyboard with DALL-E 3.
-    Focus on cinematic composition, dramatic lighting, depth of field, rich textures, and atmospheric effects.
-    Output only the new prompt string.
-    """
+def create_visual_regeneration_prompt(original_image_prompt_text, user_feedback_on_visuals, scene_data, character_definitions=None, global_style_reference=""):
+    """
+    Prompt for Gemini to rewrite an existing DALL-E image prompt based on user feedback and scene context.
+    """
+    # Reconstruct parts of what the original prompt *might* have focused on, for context
+    scene_context_summary = (
+        f"Scene Number: {scene_data.get('scene_number', 'N/A')}. "
+        f"Emotional Beat: {scene_data.get('emotional_beat', '')}. "
+        f"Setting: {scene_data.get('setting_description', '')}. "
+        f"Action: {scene_data.get('key_action', '')}. "
+        f"Characters: {', '.join(scene_data.get('characters_involved',[]))}. "
+        f"Current Style Hint: {scene_data.get('visual_style_suggestion', '')}. "
+        f"Current Camera Hint: {scene_data.get('camera_angle_suggestion', '')}."
+    )
+    
+    character_details_for_prompt = []
+    if scene_data.get('characters_involved'):
+        for char_name_in_scene in scene_data.get('characters_involved', []):
+            char_name_clean = char_name_in_scene.strip()
+            char_lookup_key = char_name_clean.lower()
+            if character_definitions and char_lookup_key in character_definitions:
+                char_visual_desc = character_definitions[char_lookup_key]
+                character_details_for_prompt.append(f"{char_name_clean} (described as: {char_visual_desc})")
+            else:
+                character_details_for_prompt.append(char_name_clean)
+    characters_narrative = f"Relevant characters and their descriptions: {', '.join(character_details_for_prompt) if character_details_for_prompt else 'None specified'}."
+
+
+    full_prompt_for_gemini = f"""
+    You are an AI assistant specializing in refining image generation prompts for DALL-E 3.
+    The user wants to modify a visual concept for a cinematic scene.
+
+    Original Scene Context:
+    {scene_context_summary}
+    {characters_narrative}
+    Global Style Reference (if any): "{global_style_reference}"
+
+    The DALL-E 3 prompt that was used to generate the previous image was:
+    "{original_image_prompt_text}"
+
+    The user provided the following feedback on the visual generated by that prompt:
+    "{user_feedback_on_visuals}"
+
+    Your task is to generate a new, revised DALL-E 3 prompt that incorporates the user's feedback to achieve the desired visual changes.
+    The new prompt should be ultra-detailed, photorealistic, and highly cinematic.
+    It should instruct DALL-E 3 to create an image that feels like a high-quality still from a major motion picture or AAA video game.
+    Maintain the core elements of the scene (setting, characters, key action) unless the feedback explicitly asks to change them.
+    Focus on translating the user's feedback into concrete visual descriptions related to composition, lighting, color, character appearance/pose, atmosphere, etc.
+    Ensure character descriptions from the context are respected and reinforced if characters are mentioned.
+    The prompt should be a single block of text.
+
+    Output ONLY the new, revised DALL-E 3 prompt string.
+    """
+    return " ".join(full_prompt_for_gemini.split())