prompt_generator.py

import streamlit as st
import torch
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

class PromptGenerator:
    def __init__(self):
        self.model = None
        self.tokenizer = None
        self.prompt_cache = {}
        
    def load_model(self):
        """Load a lightweight text generation model"""
        if self.model is None:
            with st.spinner("Loading text-to-prompt model..."):
                try:
                    # Using BART model for better prompt enhancement
                    model_name = "facebook/bart-large-cnn"
                    
                    # Load tokenizer and model separately to avoid device issues
                    self.tokenizer = AutoTokenizer.from_pretrained(model_name)
                    
                    # Load model with optimizations for memory efficiency
                    self.model = AutoModelForSeq2SeqLM.from_pretrained(
                        model_name,
                        low_cpu_mem_usage=True,
                        torch_dtype=torch.float16
                    )
                    
                    # Explicitly move to CPU to avoid meta tensor issues
                    self.model = self.model.to("cpu")
                except Exception as e:
                    st.warning(f"Error loading model: {str(e)}. Using fallback method.")
                    # If model loading fails, we'll use a simple keyword-based approach
                    self.model = None
                    self.tokenizer = None
        
        return self.model, self.tokenizer
    
    def generate_hyper_realistic_prompt(self, transcription, aspect_ratio="16:9"):
        """Generate a hyper-realistic prompt from a transcription with cinematic quality using BART model"""
        # Check cache first
        import hashlib
        cache_key = hashlib.md5((transcription + aspect_ratio).encode()).hexdigest()
        
        if cache_key in self.prompt_cache:
            return self.prompt_cache[cache_key]
        
        # Skip empty transcriptions
        if not transcription.strip():
            return ""
            
        # For horror/suspense story, use specific visual themes
        # Analyze the content to determine the scene type
        lower_trans = transcription.lower()
        
        # Create a more meaningful base prompt from the transcription
        # Extract key visual elements and create a scene description
        visual_scene = self.extract_visual_elements(transcription)
        
        # Determine scene type based on content analysis
        scene_type = "neutral"
        if any(word in lower_trans for word in ["dark", "shadow", "fear", "horror", "scary", "afraid", "terror", "scream", "blood", "death", "evil", "monster", "ghost", "nightmare", "creepy", "spooky", "haunted", "skeleton"]):
            scene_type = "horror"
        elif any(word in lower_trans for word in ["mystery", "detective", "clue", "investigate", "secret", "discover", "reveal", "hidden", "puzzle", "solve"]):
            scene_type = "mystery"
        elif any(word in lower_trans for word in ["fantasy", "magic", "wizard", "dragon", "fairy", "enchanted", "spell", "mythical", "legend", "ancient", "kingdom"]):
            scene_type = "fantasy"
        elif any(word in lower_trans for word in ["sci-fi", "future", "space", "alien", "robot", "technology", "advanced", "starship", "planet", "galaxy"]):
            scene_type = "scifi"
            
        # Select appropriate visual elements based on scene type
        if scene_type == "horror":
            style_keywords = [
                "dark atmospheric horror scene", 
                "cinematic horror", 
                "atmospheric dread", 
                "horror movie still",
                "dark gothic scene",
                "eerie lighting",
                "suspenseful moment"
            ]
            lighting = [
                "dim lighting", 
                "shadows", 
                "moonlight", 
                "eerie glow",
                "dark atmospheric lighting"
            ]
        elif scene_type == "mystery":
            style_keywords = [
                "mysterious scene", 
                "film noir style", 
                "detective story visual", 
                "suspenseful moment",
                "enigmatic scene"
            ]
            lighting = [
                "moody lighting", 
                "dramatic shadows", 
                "low key lighting", 
                "atmospheric fog"
            ]
        elif scene_type == "fantasy":
            style_keywords = [
                "fantasy scene", 
                "magical environment", 
                "enchanted setting", 
                "mythical landscape",
                "fantasy illustration style"
            ]
            lighting = [
                "magical glow", 
                "ethereal light", 
                "golden hour", 
                "mystical atmosphere"
            ]
        elif scene_type == "scifi":
            style_keywords = [
                "futuristic scene", 
                "sci-fi environment", 
                "high-tech setting", 
                "advanced technology visual",
                "science fiction concept art"
            ]
            lighting = [
                "neon lighting", 
                "holographic glow", 
                "futuristic illumination", 
                "technological ambiance"
            ]
        else:
            style_keywords = [
                "cinematic scene", 
                "photorealistic environment", 
                "detailed setting", 
                "professional photography",
                "movie still"
            ]
            lighting = [
                "natural lighting", 
                "golden hour", 
                "soft illumination", 
                "dramatic lighting"
            ]
            
        # Select quality keywords that work well with Stable Diffusion
        quality_keywords = [
            "highly detailed", 
            "8k resolution", 
            "photorealistic", 
            "detailed textures",
            "professional photography"
        ]
        
        # Select camera keywords
        camera_keywords = [
            "shallow depth of field", 
            "cinematic composition", 
            "movie still", 
            "professional photography"
        ]
        
        # Import random for selection
        import random
        
        # Select random elements from each category
        selected_style = random.choice(style_keywords)
        selected_lighting = random.choice(lighting)
        selected_quality = ", ".join(random.sample(quality_keywords, 2))
        selected_camera = random.choice(camera_keywords)
        
        # Construct a prompt that focuses on visual elements
        enhanced_prompt = f"{visual_scene}, {selected_style}, {selected_lighting}, {selected_quality}, {selected_camera}"
        
        # Add anti-text elements
        enhanced_prompt += ", no text, no words, no writing"
        
        # Clean the prompt to remove any text-generating patterns
        enhanced_prompt = self.clean_prompt_for_image_generation(enhanced_prompt)
        
        # Cache the result
        self.prompt_cache[cache_key] = enhanced_prompt
        
        return enhanced_prompt
        
    def extract_visual_elements(self, text):
        """Extract key visual elements from text to create a scene description"""
        import re
        
        # Clean the text
        text = text.strip()
        
        # Look for visual elements using NLP patterns
        visual_elements = []
        
        # Look for locations/settings
        locations = re.findall(r'(inside|outside|in the|at the|near the|by the|on the) ([a-z ]+)', text.lower())
        for loc in locations:
            if len(loc[1]) > 3:  # Avoid very short words
                visual_elements.append(f"{loc[0]} {loc[1]}")
                
        # Look for objects
        objects = re.findall(r'the ([a-z]+)', text.lower())
        for obj in objects:
            if len(obj) > 3 and obj not in ["that", "this", "then", "than", "they", "them", "with", "from", "were", "when", "what", "which"]:
                visual_elements.append(obj)
                
        # Look for adjectives followed by nouns
        adj_nouns = re.findall(r'([a-z]+) ([a-z]+)', text.lower())
        for adj_noun in adj_nouns:
            if len(adj_noun[0]) > 2 and len(adj_noun[1]) > 3:
                visual_elements.append(f"{adj_noun[0]} {adj_noun[1]}")
        
        # If we found visual elements, use them
        if visual_elements:
            # Take up to 3 unique elements
            unique_elements = list(set(visual_elements))
            selected_elements = unique_elements[:3]
            return ", ".join(selected_elements)
        
        # If no visual elements found, use the first part of the text
        if len(text) > 100:
            return text[:100]
        return text
        
    def clean_prompt_for_image_generation(self, prompt):
        """Clean prompt to avoid patterns that might cause text rendering in images"""
        # Remove patterns that might cause text rendering
        import re
        
        # Remove explicit text formatting instructions
        cleaned = re.sub(r'text\s+that\s+says', '', prompt, flags=re.IGNORECASE)
        cleaned = re.sub(r'with\s+text', '', cleaned, flags=re.IGNORECASE)
        cleaned = re.sub(r'showing\s+text', '', cleaned, flags=re.IGNORECASE)
        cleaned = re.sub(r'displaying\s+text', '', cleaned, flags=re.IGNORECASE)
        cleaned = re.sub(r'with\s+the\s+words', '', cleaned, flags=re.IGNORECASE)
        cleaned = re.sub(r'caption', '', cleaned, flags=re.IGNORECASE)
        cleaned = re.sub(r'title', '', cleaned, flags=re.IGNORECASE)
        
        # Remove quotes which might encourage text
        cleaned = re.sub(r'["\'].*?["\']', '', cleaned)
        
        return cleaned
    
    def generate_optimized_prompt(self, transcription, aspect_ratio="16:9"):
        """Generate an optimized prompt from a single transcription"""
        # This is now a wrapper for the hyper-realistic prompt generator
        return self.generate_hyper_realistic_prompt(transcription, aspect_ratio)
    
    def generate_prompts(self, text, num_segments=5, aspect_ratio="16:9"):
        """Generate image prompts from the transcription"""
        # Split text into segments
        words = text.split()
        segment_size = max(1, len(words) // num_segments)
        segments = []
        
        for i in range(0, len(words), segment_size):
            segment = " ".join(words[i:i+segment_size])
            segments.append(segment)
        
        # Limit to the desired number of segments
        segments = segments[:num_segments]
        
        # Generate a creative prompt for each segment
        prompts = []
        for segment in segments:
            # Create an enhanced prompt
            enhanced_prompt = self.generate_hyper_realistic_prompt(segment, aspect_ratio)
            prompts.append(enhanced_prompt)
        
        return prompts, segments
    
    def generate_optimized_prompts(self, transcriptions, parallel=False, max_workers=4, aspect_ratio="16:9"):
        """Generate optimized prompts from transcribed segments with parallel processing"""
        import concurrent.futures
        
        if parallel and len(transcriptions) > 1:
            # Process in parallel
            with concurrent.futures.ThreadPoolExecutor(max_workers=max_workers) as executor:
                # Create a function that includes aspect ratio
                def generate_with_aspect(trans):
                    return self.generate_hyper_realistic_prompt(trans, aspect_ratio)
                
                # Map with the new function
                prompts = list(executor.map(generate_with_aspect, transcriptions))
        else:
            # Process sequentially
            prompts = [self.generate_hyper_realistic_prompt(trans, aspect_ratio) for trans in transcriptions]
            
        return prompts
    
    def clear_cache(self):
        """Clear the prompt cache"""
        self.prompt_cache = {}
        return True