import gradio as gr from gradio_pdf import PDF import numpy as np import random import torch import spaces import math import os import yaml import io import tempfile import shutil import uuid import time import json from typing import List, Tuple, Dict, Optional from datetime import datetime, timedelta from pathlib import Path from PIL import Image from diffusers import DiffusionPipeline, FlowMatchEulerDiscreteScheduler from huggingface_hub import InferenceClient from reportlab.lib.pagesizes import A4 from reportlab.pdfgen import canvas from reportlab.pdfbase import pdfmetrics from reportlab.lib.utils import ImageReader from PyPDF2 import PdfReader, PdfWriter # --- Style Presets Loading --- def load_style_presets(): """Load style presets from YAML file.""" try: with open('style_presets.yaml', 'r') as f: data = yaml.safe_load(f) # Filter only enabled presets presets = {k: v for k, v in data['presets'].items() if v.get('enabled', True)} return presets except Exception as e: print(f"Error loading style presets: {e}") return {"no_style": {"id": "no_style", "label": "No style (custom)", "prompt_prefix": "", "prompt_suffix": "", "negative_prompt": ""}} # Load presets at startup STYLE_PRESETS = load_style_presets() # --- Page Layouts Loading --- def load_page_layouts(): """Load page layouts from YAML file.""" try: with open('page_layouts.yaml', 'r') as f: data = yaml.safe_load(f) return data['layouts'] except Exception as e: print(f"Error loading page layouts: {e}") # Fallback to basic layouts return { "1_image": [{"id": "full_page", "label": "Full Page", "positions": [[0.05, 0.05, 0.9, 0.9]]}], "2_images": [{"id": "horizontal_split", "label": "Horizontal Split", "positions": [[0.05, 0.05, 0.425, 0.9], [0.525, 0.05, 0.425, 0.9]]}], "3_images": [{"id": "grid", "label": "Grid", "positions": [[0.05, 0.05, 0.283, 0.5], [0.358, 0.05, 0.283, 0.5], [0.666, 0.05, 0.283, 0.5]]}], "4_images": [{"id": "grid_2x2", "label": "2x2 Grid", "positions": [[0.05, 0.05, 0.425, 0.425], [0.525, 0.05, 0.425, 0.425], [0.05, 0.525, 0.425, 0.425], [0.525, 0.525, 0.425, 0.425]]}] } # Load layouts at startup PAGE_LAYOUTS = load_page_layouts() def get_layout_choices(num_images: int) -> List[Tuple[str, str]]: """Get available layout choices for a given number of images.""" key = f"{num_images}_image" if num_images == 1 else f"{num_images}_images" if key in PAGE_LAYOUTS: return [(layout["label"], layout["id"]) for layout in PAGE_LAYOUTS[key]] # Return empty list if no layouts found (shouldn't happen with our config) return [("Default", "default")] def get_layout_metadata(layout_id: str, num_images: int) -> List[Dict]: """Get metadata for each panel in a layout. Args: layout_id: ID of the selected layout num_images: Total number of images Returns: List of metadata dicts with panel_type, focus, composition, shot_type, and camera_angle """ key = f"{num_images}_image" if num_images == 1 else f"{num_images}_images" layouts = PAGE_LAYOUTS.get(key, []) layout = next((l for l in layouts if l["id"] == layout_id), None) if layout and "metadata" in layout: return layout["metadata"] # Fallback metadata if not found fallback_meta = { "panel_type": "action", "focus": "character", "composition": "square", "shot_type": "medium", "camera_angle": "eye_level" } return [fallback_meta] * num_images def get_random_style_preset(): """Get a random style preset (excluding 'no_style' and 'random').""" eligible_keys = [k for k in STYLE_PRESETS.keys() if k not in ['no_style', 'random']] if eligible_keys: return random.choice(eligible_keys) return 'no_style' def apply_style_preset(prompt, style_preset_key, custom_style_text=""): """ Apply style preset to the prompt. Args: prompt: The user's base prompt style_preset_key: The key of the selected style preset custom_style_text: Custom style text when 'no_style' is selected Returns: tuple: (styled_prompt, negative_prompt) """ if style_preset_key == 'no_style': # Use custom style text if provided if custom_style_text and custom_style_text.strip(): styled_prompt = f"{custom_style_text}, {prompt}" else: styled_prompt = prompt return styled_prompt, "" if style_preset_key == 'random': # Select a random style style_preset_key = get_random_style_preset() if style_preset_key in STYLE_PRESETS: preset = STYLE_PRESETS[style_preset_key] prefix = preset.get('prompt_prefix', '') suffix = preset.get('prompt_suffix', '') negative = preset.get('negative_prompt', '') # Build the styled prompt parts = [] if prefix: parts.append(prefix) parts.append(prompt) if suffix: parts.append(suffix) styled_prompt = ', '.join(parts) return styled_prompt, negative # Fallback to original prompt if preset not found return prompt, "" # --- Story Generation using Hugging Face InferenceClient --- def generate_story_scenes(story_prompt, num_scenes, style_context="", panel_metadata=None): """ Generates a sequence of scene descriptions with captions and dialogues. Args: story_prompt: The user's story prompt num_scenes: Number of scenes to generate style_context: Optional style context to consider panel_metadata: List of metadata dicts for each panel Returns: List of dicts with 'caption' and 'dialogue' keys """ # Ensure HF_TOKEN is set api_key = os.environ.get("HF_TOKEN") if not api_key: print("HF_TOKEN not set, using fallback scene generation") # Simple fallback - just split the prompt into scenes fallback_scenes = [] for i in range(num_scenes): fallback_scenes.append({ "caption": f"{story_prompt} (scene {i+1} of {num_scenes})", "dialogue": "" }) return fallback_scenes # Initialize the client client = InferenceClient( provider="cerebras", api_key=api_key, ) # Build panel descriptions from metadata panel_descriptions = [] if panel_metadata and len(panel_metadata) == num_scenes: for i, meta in enumerate(panel_metadata): panel_type = meta.get('panel_type', 'action') focus = meta.get('focus', 'character') composition = meta.get('composition', 'square') shot_type = meta.get('shot_type', 'medium') camera_angle = meta.get('camera_angle', 'eye_level') # Format shot type for readability shot_display = shot_type.replace('_', ' ').title() angle_display = camera_angle.replace('_', ' ').title() # Create a descriptive text for this panel desc = f"Panel {i+1}/{num_scenes} - {composition.upper()} composition, {shot_display} shot at {angle_display} angle, {panel_type} panel focusing on {focus}" panel_descriptions.append(desc) else: # Fallback if no metadata panel_descriptions = [f"Panel {i+1}/{num_scenes}" for i in range(num_scenes)] # Create system prompt with panel-specific guidance system_prompt = f"""You are a story writer with expertise in cinematography and visual storytelling. Generate exactly {num_scenes} panels for a short story, based on the user's story prompt. PANEL LAYOUT INFORMATION: The page has {num_scenes} panels with the following characteristics: {chr(10).join(f"- {desc}" for desc in panel_descriptions)} IMPORTANT INSTRUCTIONS: 1. Output ONLY a YAML list with exactly {num_scenes} items 2. Each item must have exactly two fields: - caption: A detailed visual description of the scene (describe characters, clothing, location, action, expressions) - dialogue: Natural language description of what the character says/exclaims/shouts (can be empty string if no dialogue) 3. **ADAPT EACH SCENE TO ITS PANEL TYPE:** - ESTABLISHING panels: Describe the full environment, setting, atmosphere, time of day, location details - ACTION panels: Focus on dynamic movement, motion lines, impact, energy, physical activity - CLOSEUP panels: Describe facial features, eyes, expressions, emotions in extreme detail - DIALOGUE panels: Focus on character interactions, body language during conversation - REACTION panels: Emphasize emotional responses, facial expressions, body language - DETAIL panels: Zoom in on specific objects, hands, symbols, or small but important elements - TRANSITION panels: Show passage of time, change of location, or connecting moments - SPLASH panels: Epic, dramatic, full-scene moments with maximum visual impact 4. **ADAPT TO SHOT TYPE (CINEMATOGRAPHY):** - EXTREME WIDE SHOT: Vast landscapes, tiny characters in massive environments, epic scale - WIDE SHOT: Full scene with characters and environment, establishing context - FULL SHOT: Entire character from head to toe, showing their full body and stance - MEDIUM FULL SHOT: Character from knees up, showing most of body with some environment - MEDIUM SHOT: Character from waist up, balanced between character and setting - MEDIUM CLOSEUP: Head and shoulders, focusing on face while showing some context - CLOSEUP: Face filling frame, detailed facial features and expressions - EXTREME CLOSEUP: Tiny detail - just eyes, hands, mouth, or specific object filling frame 5. **ADAPT TO CAMERA ANGLE:** - EYE LEVEL: Neutral, straightforward angle - camera at character's eye level - HIGH ANGLE: Camera looking down on subject - can make them seem vulnerable, small, or overwhelmed - LOW ANGLE: Camera looking up at subject - makes them seem powerful, imposing, heroic - OVERHEAD/BIRD'S EYE: Camera directly above looking down - shows spatial relationships, isolation - DUTCH ANGLE/CANTED: Tilted camera - creates tension, disorientation, chaos, unease - OVER THE SHOULDER (OTS): Camera behind one character looking at another - intimate conversation - POV (Point of View): Camera as character's eyes - immersive, first-person perspective 6. **ADAPT TO COMPOSITION:** - WIDE/LANDSCAPE: Emphasize horizontal elements, panoramic views, sweeping scenes, breadth - TALL/PORTRAIT: Emphasize vertical elements, full-body shots, top-to-bottom action, height - SQUARE: Balanced composition, centered subjects, symmetrical arrangements 7. **ADAPT TO FOCUS:** - CHARACTER: Detailed character description (appearance, clothing, pose, expression) - CHARACTERS (plural): Multiple people, their relationships, positioning, interactions - ENVIRONMENT: Setting details, location, atmosphere, background elements, mood - EVENT: What's happening, the action, the moment being captured, the incident - EMOTION: Facial expression, body language, emotional state, feelings - OBJECT: Detailed description of an important item, prop, symbol, or artifact - ACTION: Movement, impact, dynamic poses, energy, motion 8. **CONSIDER PANEL PROGRESSION:** - You're creating panel X of {num_scenes} - consider where you are in the story flow - Early panels (1-2/{num_scenes}): Establish setting and introduce characters - Middle panels: Build action, develop conflict, show character reactions - Later panels ({num_scenes-1}-{num_scenes}/{num_scenes}): Resolve the moment, provide reaction or cliffhanger 9. For captions: Be VERY descriptive. Include shot type language like "wide shot of...", "close-up on...", "overhead view of...". Repeat character descriptions in each scene if needed. 10. For dialogue: Write as natural language action: "The [character] says: [what they say]" or "The [character] exclaims: [what they exclaim]" - DO NOT include character names in the dialogue text itself - Use verbs like: says, exclaims, shouts, whispers, asks, replies, thinks, mutters, screams 11. Keep continuity between scenes to tell a coherent story 12. Make each scene visually distinct but connected to the narrative Example output format: - caption: "Extreme wide shot from high angle of a dark alley at night, rain pouring down heavily, neon signs casting red and blue reflections in vast puddles covering the ground, tall buildings looming menacingly on both sides creating a narrow canyon, a young woman with long red hair wearing a blue detective coat stands small in the center of the frame examining glowing footprints on the wet pavement with a magnifying glass, dramatic lighting from above" dialogue: "The detective whispers to herself: These tracks... they're not human" - caption: "Extreme close-up at eye level of the detective's piercing green eyes widening in shock and fear, her pupils dilating rapidly, individual beads of rain clinging to her dark eyelashes, her face illuminated by an eerie pulsing blue glow from below, wrinkles forming on her forehead" dialogue: "" - caption: "Full shot at low angle, the red-haired detective in the blue coat leaping backwards dynamically with motion blur streaks, her coat billowing dramatically, a massive jagged shark fin erupting violently from a puddle behind her, water exploding upward in huge spray with droplets frozen mid-air, her expression one of pure terror, arms flailing" dialogue: "The detective shouts at the top of her lungs: OH NO! SHARKS IN THE CITY!" Generate exactly {num_scenes} scenes. Output ONLY the YAML list, no other text.""" # Format the messages messages = [ {"role": "system", "content": system_prompt}, {"role": "user", "content": f"Create {num_scenes} scenes for this story: {story_prompt}"} ] try: # Call the API completion = client.chat.completions.create( model="Qwen/Qwen3-235B-A22B-Instruct-2507", messages=messages, temperature=0.7, max_tokens=2000, ) response = completion.choices[0].message.content # Parse the YAML response scenes = parse_yaml_scenes(response, num_scenes) return scenes except Exception as e: print(f"Error during story generation: {e}") # Fallback to simple scene splitting fallback_scenes = [] for i in range(num_scenes): fallback_scenes.append({ "caption": f"{story_prompt} (part {i+1} of {num_scenes})", "dialogue": "" }) return fallback_scenes def parse_yaml_scenes(yaml_text, expected_count): """ Parse YAML text to extract scene captions and dialogues. """ try: # Clean up the text - remove markdown code blocks if present yaml_text = yaml_text.strip() if yaml_text.startswith("```yaml"): yaml_text = yaml_text[7:] if yaml_text.startswith("```"): yaml_text = yaml_text[3:] if yaml_text.endswith("```"): yaml_text = yaml_text[:-3] # Parse YAML scenes = yaml.safe_load(yaml_text) if not isinstance(scenes, list): raise ValueError("Expected a list of scenes") # Validate and clean scenes valid_scenes = [] for scene in scenes: if isinstance(scene, dict) and 'caption' in scene: valid_scenes.append({ 'caption': str(scene.get('caption', '')), 'dialogue': str(scene.get('dialogue', '')) }) # Ensure we have the expected number of scenes while len(valid_scenes) < expected_count: valid_scenes.append({ 'caption': 'continuation of the story', 'dialogue': '' }) return valid_scenes[:expected_count] except Exception as e: print(f"Error parsing YAML scenes: {e}") # Return fallback scenes return [{'caption': 'scene description', 'dialogue': ''} for _ in range(expected_count)] def get_caption_language(prompt): """Detects if the prompt contains Chinese characters.""" ranges = [ ('\u4e00', '\u9fff'), # CJK Unified Ideographs ] for char in prompt: if any(start <= char <= end for start, end in ranges): return 'zh' return 'en' # --- Model Loading --- # Use the new lightning-fast model setup ckpt_id = "Qwen/Qwen-Image" # Scheduler configuration from the Qwen-Image-Lightning repository scheduler_config = { "base_image_seq_len": 256, "base_shift": math.log(3), "invert_sigmas": False, "max_image_seq_len": 8192, "max_shift": math.log(3), "num_train_timesteps": 1000, "shift": 1.0, "shift_terminal": None, "stochastic_sampling": False, "time_shift_type": "exponential", "use_beta_sigmas": False, "use_dynamic_shifting": True, "use_exponential_sigmas": False, "use_karras_sigmas": False, } scheduler = FlowMatchEulerDiscreteScheduler.from_config(scheduler_config) pipe = DiffusionPipeline.from_pretrained( ckpt_id, scheduler=scheduler, torch_dtype=torch.bfloat16 ).to("cuda") # Load LoRA weights for acceleration pipe.load_lora_weights( "lightx2v/Qwen-Image-Lightning", weight_name="Qwen-Image-Lightning-8steps-V1.1.safetensors" ) pipe.fuse_lora() # --- UI Constants and Helpers --- MAX_SEED = np.iinfo(np.int32).max def get_image_size_for_position(position_data, image_index, num_images, max_resolution=1024): """ Calculate EXACT image dimensions to match layout aspect ratio perfectly. This function calculates pixel dimensions that precisely match the aspect ratio of the layout rectangle to ensure images fill their containers without floating. Args: position_data: Layout position data [x, y, width, height] in relative units (0-1) image_index: Index of the current image (0-based) num_images: Total number of images in the layout max_resolution: Maximum resolution for any dimension (default 1024) Returns: tuple: (width, height) with exact aspect ratio matching the layout """ if not position_data: return max_resolution, max_resolution # Default square x_rel, y_rel, w_rel, h_rel = position_data # Calculate the EXACT aspect ratio from the layout rectangle # This is crucial - we must match this aspect ratio precisely # A4 page dimensions in points (must match PDF generation) page_width, page_height = 595.27, 841.89 # Account for the page's aspect ratio when calculating layout aspect ratio layout_aspect_ratio = (w_rel / h_rel) * (page_width / page_height) if h_rel > 0 else 1.0 # Scale to max_resolution while maintaining EXACT aspect ratio if layout_aspect_ratio >= 1: # Wider than tall width = max_resolution height = max_resolution / layout_aspect_ratio else: # Taller than wide height = max_resolution width = max_resolution * layout_aspect_ratio # Round to nearest 8 pixels for model compatibility # Using 8px instead of 64px preserves aspect ratio much better # Most diffusion models work well with multiples of 8 width = round(width / 8) * 8 height = round(height / 8) * 8 # After rounding, ensure we maintain the aspect ratio as closely as possible # and don't exceed max_resolution if width > max_resolution: width = max_resolution height = round((max_resolution / layout_aspect_ratio) / 8) * 8 if height > max_resolution: height = max_resolution width = round((max_resolution * layout_aspect_ratio) / 8) * 8 # Ensure minimum size of 256px (reduced from 384 for more flexibility) # while maintaining the layout aspect ratio min_size = 256 if width < min_size or height < min_size: if layout_aspect_ratio >= 1: # Wider image width = max(min_size, width) height = round((width / layout_aspect_ratio) / 8) * 8 else: # Taller image height = max(min_size, height) width = round((height * layout_aspect_ratio) / 8) * 8 # Final safety checks width = max(min_size, min(int(width), max_resolution)) height = max(min_size, min(int(height), max_resolution)) return width, height def get_layout_position_for_image(layout_id, num_images, image_index): """Get the position data for a specific image in a layout. Args: layout_id: ID of the selected layout num_images: Total number of images image_index: Index of the current image (0-based) Returns: Position data [x, y, width, height] or None """ key = f"{num_images}_image" if num_images == 1 else f"{num_images}_images" layouts = PAGE_LAYOUTS.get(key, []) layout = next((l for l in layouts if l["id"] == layout_id), None) if layout and "positions" in layout: positions = layout["positions"] if image_index < len(positions): return positions[image_index] # Fallback positions for each number of images fallback_positions = { 1: [[0.05, 0.05, 0.9, 0.9]], 2: [[0.05, 0.05, 0.425, 0.9], [0.525, 0.05, 0.425, 0.9]], 3: [[0.05, 0.25, 0.283, 0.5], [0.358, 0.25, 0.283, 0.5], [0.666, 0.25, 0.283, 0.5]], 4: [[0.05, 0.05, 0.425, 0.425], [0.525, 0.05, 0.425, 0.425], [0.05, 0.525, 0.425, 0.425], [0.525, 0.525, 0.425, 0.425]], 5: [[0.05, 0.05, 0.9, 0.3], [0.05, 0.4, 0.283, 0.55], [0.358, 0.4, 0.283, 0.55], [0.666, 0.4, 0.283, 0.275], [0.666, 0.7, 0.283, 0.275]], 6: [[0.05, 0.05, 0.425, 0.283], [0.525, 0.05, 0.425, 0.283], [0.05, 0.358, 0.425, 0.283], [0.525, 0.358, 0.425, 0.283], [0.05, 0.666, 0.425, 0.283], [0.525, 0.666, 0.425, 0.283]], 7: [[0.28, 0.02, 0.44, 0.3], [0.02, 0.25, 0.3, 0.25], [0.68, 0.25, 0.3, 0.25], [0.25, 0.35, 0.5, 0.3], [0.02, 0.52, 0.3, 0.25], [0.68, 0.52, 0.3, 0.25], [0.28, 0.68, 0.44, 0.3]], 8: [[0.02, 0.02, 0.23, 0.47], [0.27, 0.02, 0.23, 0.47], [0.52, 0.02, 0.23, 0.47], [0.77, 0.02, 0.21, 0.47], [0.02, 0.51, 0.23, 0.47], [0.27, 0.51, 0.23, 0.47], [0.52, 0.51, 0.23, 0.47], [0.77, 0.51, 0.21, 0.47]] } positions = fallback_positions.get(num_images, fallback_positions[1]) if image_index < len(positions): return positions[image_index] return [0.05, 0.05, 0.9, 0.9] # Ultimate default # --- Session Management Functions --- class SessionManager: """Manages user session data and temporary file storage.""" def __init__(self, session_id: str = None): self.session_id = session_id or str(uuid.uuid4()) self.base_dir = Path(tempfile.gettempdir()) / "gradio_comic_sessions" self.session_dir = self.base_dir / self.session_id self.session_dir.mkdir(parents=True, exist_ok=True) self.metadata_file = self.session_dir / "metadata.json" self.pdf_path = self.session_dir / "comic.pdf" self.load_or_create_metadata() def load_or_create_metadata(self): """Load existing metadata or create new.""" if self.metadata_file.exists(): with open(self.metadata_file, 'r') as f: self.metadata = json.load(f) else: self.metadata = { "created_at": datetime.now().isoformat(), "pages": [], "total_pages": 0 } self.save_metadata() def save_metadata(self): """Save metadata to file.""" with open(self.metadata_file, 'w') as f: json.dump(self.metadata, f, indent=2) def add_page(self, images: List[Image.Image], layout_id: str, seeds: List[int]): """Add a new page to the session.""" page_num = self.metadata["total_pages"] + 1 page_dir = self.session_dir / f"page_{page_num}" page_dir.mkdir(exist_ok=True) # Save images image_paths = [] for i, img in enumerate(images): img_path = page_dir / f"image_{i+1}.jpg" img.save(img_path, 'JPEG', quality=95) image_paths.append(str(img_path)) # Update metadata self.metadata["pages"].append({ "page_num": page_num, "layout_id": layout_id, "num_images": len(images), "image_paths": image_paths, "seeds": seeds, "created_at": datetime.now().isoformat() }) self.metadata["total_pages"] = page_num self.save_metadata() return page_num def get_all_pages_images(self) -> List[Tuple[List[Image.Image], str, int]]: """Get all images from all pages.""" pages_data = [] for page in self.metadata["pages"]: images = [] for img_path in page["image_paths"]: if Path(img_path).exists(): images.append(Image.open(img_path)) if images: pages_data.append((images, page["layout_id"], page["num_images"])) return pages_data def cleanup_old_sessions(self, max_age_hours: int = 24): """Clean up sessions older than max_age_hours.""" if not self.base_dir.exists(): return cutoff_time = datetime.now() - timedelta(hours=max_age_hours) for session_dir in self.base_dir.iterdir(): if session_dir.is_dir(): metadata_file = session_dir / "metadata.json" if metadata_file.exists(): try: with open(metadata_file, 'r') as f: metadata = json.load(f) created_at = datetime.fromisoformat(metadata["created_at"]) if created_at < cutoff_time: shutil.rmtree(session_dir) print(f"Cleaned up old session: {session_dir.name}") except Exception as e: print(f"Error cleaning session {session_dir.name}: {e}") # --- PDF Generation Functions --- def create_single_page_pdf(images: List[Image.Image], layout_id: str, num_images: int) -> bytes: """ Create a single PDF page with images arranged according to the selected layout. Args: images: List of PIL images layout_id: ID of the selected layout num_images: Number of images to include Returns: PDF page as bytes """ # Create a bytes buffer for the PDF pdf_buffer = io.BytesIO() # Create canvas with A4 size pdf = canvas.Canvas(pdf_buffer, pagesize=A4) page_width, page_height = A4 # Get the layout configuration key = f"{num_images}_image" if num_images == 1 else f"{num_images}_images" layouts = PAGE_LAYOUTS.get(key, []) layout = next((l for l in layouts if l["id"] == layout_id), None) if not layout: # Fallback to default grid layout with proper spacing if num_images == 1: positions = [[0.02, 0.02, 0.96, 0.96]] elif num_images == 2: positions = [[0.02, 0.02, 0.47, 0.96], [0.51, 0.02, 0.47, 0.96]] elif num_images == 3: positions = [[0.02, 0.2, 0.31, 0.6], [0.345, 0.2, 0.31, 0.6], [0.67, 0.2, 0.31, 0.6]] elif num_images == 4: positions = [[0.02, 0.02, 0.47, 0.47], [0.51, 0.02, 0.47, 0.47], [0.02, 0.51, 0.47, 0.47], [0.51, 0.51, 0.47, 0.47]] elif num_images == 5: positions = [[0.02, 0.02, 0.96, 0.44], [0.02, 0.48, 0.31, 0.5], [0.345, 0.48, 0.31, 0.5], [0.67, 0.48, 0.31, 0.24], [0.67, 0.74, 0.31, 0.24]] elif num_images == 6: positions = [[0.02, 0.02, 0.47, 0.31], [0.51, 0.02, 0.47, 0.31], [0.02, 0.345, 0.47, 0.31], [0.51, 0.345, 0.47, 0.31], [0.02, 0.67, 0.47, 0.31], [0.51, 0.67, 0.47, 0.31]] elif num_images == 7: positions = [[0.28, 0.02, 0.44, 0.3], [0.02, 0.25, 0.3, 0.25], [0.68, 0.25, 0.3, 0.25], [0.25, 0.35, 0.5, 0.3], [0.02, 0.52, 0.3, 0.25], [0.68, 0.52, 0.3, 0.25], [0.28, 0.68, 0.44, 0.3]] elif num_images == 8: positions = [[0.02, 0.02, 0.23, 0.47], [0.27, 0.02, 0.23, 0.47], [0.52, 0.02, 0.23, 0.47], [0.77, 0.02, 0.21, 0.47], [0.02, 0.51, 0.23, 0.47], [0.27, 0.51, 0.23, 0.47], [0.52, 0.51, 0.23, 0.47], [0.77, 0.51, 0.21, 0.47]] else: positions = [[0.02, 0.02, 0.96, 0.96]] else: positions = layout["positions"] # Draw each image according to the layout for i, (image, pos) in enumerate(zip(images[:num_images], positions)): if i >= len(images): break x_rel, y_rel, w_rel, h_rel = pos # Add small padding between panels (1% of page dimensions) padding = 0.01 # Apply padding to prevent images from touching edges if x_rel < padding: x_rel = padding if y_rel < padding: y_rel = padding if x_rel + w_rel > 1 - padding: w_rel = 1 - padding - x_rel if y_rel + h_rel > 1 - padding: h_rel = 1 - padding - y_rel # Convert relative positions to absolute positions # Note: In ReportLab, y=0 is at the bottom x = x_rel * page_width y = (1 - y_rel - h_rel) * page_height # Flip Y coordinate width = w_rel * page_width height = h_rel * page_height # Calculate aspect ratios for comparison img_aspect = image.width / image.height layout_aspect = width / height aspect_diff = abs(img_aspect - layout_aspect) / layout_aspect # If aspect ratios match closely (within 2%), fill the space completely # Otherwise, preserve aspect ratio to avoid distortion if aspect_diff < 0.02: # Less than 2% difference # Aspect ratios match well - fill the space completely actual_width = width actual_height = height actual_x = x actual_y = y else: # Significant aspect ratio difference - preserve it to avoid distortion if img_aspect > layout_aspect: # Image is wider than the layout space new_height = width / img_aspect y_offset = (height - new_height) / 2 actual_width = width actual_height = new_height actual_x = x actual_y = y + y_offset else: # Image is taller than the layout space new_width = height * img_aspect x_offset = (width - new_width) / 2 actual_width = new_width actual_height = height actual_x = x + x_offset actual_y = y # Convert PIL image to format suitable for ReportLab img_buffer = io.BytesIO() image.save(img_buffer, format='JPEG', quality=95) img_buffer.seek(0) # Draw the image on the PDF # When aspect ratios match (aspect_diff < 0.02), we fill completely # Otherwise we preserve aspect ratio to prevent distortion pdf.drawImage(ImageReader(img_buffer), actual_x, actual_y, width=actual_width, height=actual_height, preserveAspectRatio=(aspect_diff >= 0.02), mask='auto') # Save the PDF pdf.save() # Get the PDF bytes pdf_buffer.seek(0) pdf_bytes = pdf_buffer.read() return pdf_bytes def create_multi_page_pdf(session_manager: SessionManager) -> str: """ Create a multi-page PDF from all pages in the session. Args: session_manager: SessionManager instance with page data Returns: Path to the created PDF file """ pages_data = session_manager.get_all_pages_images() if not pages_data: return None # Create PDF writer pdf_writer = PdfWriter() # Create each page for images, layout_id, num_images in pages_data: page_pdf_bytes = create_single_page_pdf(images, layout_id, num_images) # Read the single page PDF page_pdf_reader = PdfReader(io.BytesIO(page_pdf_bytes)) # Add the page to the writer for page in page_pdf_reader.pages: pdf_writer.add_page(page) # Write to file with explicit flushing pdf_path = session_manager.pdf_path with open(pdf_path, 'wb') as f: pdf_writer.write(f) f.flush() # Flush Python's internal buffer os.fsync(f.fileno()) # Ensure OS writes to disk # Small delay to ensure file system catches up time.sleep(0.1) return str(pdf_path) # --- Main Inference Function (with session support) --- @spaces.GPU(duration=240) # Increased duration for up to 8 images def infer_page( prompt, guidance_scale=1.0, num_inference_steps=8, style_preset="no_style", custom_style_text="", num_images=1, layout="default", max_resolution=1024, session_state=None, progress=gr.Progress(track_tqdm=True), ): """ Generates images for a new page and adds them to the PDF. Args: prompt (str): The text prompt to generate images from. guidance_scale (float): Corresponds to `true_cfg_scale`. num_inference_steps (int): The number of denoising steps. style_preset (str): The key of the style preset to apply. custom_style_text (str): Custom style text when 'no_style' is selected. num_images (int): Number of images to generate (1-8). layout (str): The layout ID for arranging images in the PDF. max_resolution: Maximum resolution for any dimension. session_state: Current session state dictionary. progress (gr.Progress): A Gradio Progress object to track generation. Returns: tuple: Updated session state, PDF path, and updated button label. """ # Initialize or retrieve session if session_state is None or "session_id" not in session_state: session_state = {"session_id": str(uuid.uuid4()), "page_count": 0} session_manager = SessionManager(session_state["session_id"]) # Clean up old sessions periodically if random.random() < 0.1: # 10% chance to cleanup on each request session_manager.cleanup_old_sessions() # Check page limit (reduced to 24 for performance) if session_manager.metadata["total_pages"] >= 24: return session_state, None, f"Page limit reached" generated_images = [] used_seeds = [] # Get panel metadata for this layout panel_metadata = get_layout_metadata(layout, int(num_images)) # Debug: print metadata print(f"\n=== LAYOUT METADATA for {layout} with {num_images} images ===") for i, meta in enumerate(panel_metadata): shot_type = meta.get('shot_type', 'medium').replace('_', ' ').title() camera_angle = meta.get('camera_angle', 'eye_level').replace('_', ' ').title() print(f"Panel {i+1}/{num_images}: {meta['panel_type']} | Focus: {meta['focus']} | {meta['composition']} | {shot_type} @ {camera_angle}") print("=" * 80 + "\n") # Generate story scenes with metadata progress(0, f"Generating story with {num_images} scenes...") scenes = generate_story_scenes(prompt, int(num_images), style_preset, panel_metadata) # Generate the requested number of images for i in range(int(num_images)): progress((i + 0.5) / num_images, f"Generating image {i+1} of {num_images} for page {session_manager.metadata['total_pages'] + 1}") current_seed = random.randint(0, MAX_SEED) # Always randomize seed # Get optimal aspect ratio based on position in layout position_data = get_layout_position_for_image(layout, int(num_images), i) # Use scene caption and dialogue for this image scene_prompt = scenes[i]['caption'] scene_dialogue = scenes[i]['dialogue'] # Get metadata for this panel panel_meta = panel_metadata[i] if i < len(panel_metadata) else {} # Debug output print(f"\n--- Generating Panel {i+1}/{num_images} ---") print(f"Type: {panel_meta.get('panel_type', 'unknown')}") print(f"Focus: {panel_meta.get('focus', 'unknown')}") print(f"Composition: {panel_meta.get('composition', 'unknown')}") shot_display = panel_meta.get('shot_type', 'medium').replace('_', ' ').title() angle_display = panel_meta.get('camera_angle', 'eye_level').replace('_', ' ').title() print(f"Shot: {shot_display} at {angle_display} angle") print(f"Caption: {scene_prompt[:100]}..." if len(scene_prompt) > 100 else f"Caption: {scene_prompt}") print(f"Dialogue: {scene_dialogue if scene_dialogue else '(none)'}") print("-" * 80) # Generate single image with automatic aspect ratio image, used_seed = infer_single_auto( prompt=scene_prompt, seed=current_seed, randomize_seed=False, # We handle randomization here position_data=position_data, image_index=i, num_images=int(num_images), guidance_scale=guidance_scale, num_inference_steps=num_inference_steps, dialogue=scene_dialogue, # Pass dialogue separately style_preset=style_preset, custom_style_text=custom_style_text, max_resolution=max_resolution, ) generated_images.append(image) used_seeds.append(used_seed) # Add page to session progress(0.8, "Adding page to document...") page_num = session_manager.add_page(generated_images, layout, used_seeds) # Create multi-page PDF progress(0.9, "Creating PDF...") pdf_path = create_multi_page_pdf(session_manager) progress(1.0, "Done!") # Update session state session_state["page_count"] = page_num # Next button label next_page_num = page_num + 1 button_label = f"Generate page {next_page_num}" if next_page_num <= 24 else "Page limit reached" return session_state, pdf_path, button_label # New inference function with automatic aspect ratio def infer_single_auto( prompt, seed=42, randomize_seed=False, position_data=None, image_index=0, num_images=1, guidance_scale=1.0, num_inference_steps=8, dialogue="", style_preset="no_style", custom_style_text="", max_resolution=1024, ): """ Generates an image with automatically determined aspect ratio based on layout position. """ if randomize_seed: seed = random.randint(0, MAX_SEED) # Automatically determine image size based on position with custom max resolution width, height = get_image_size_for_position(position_data, image_index, num_images, max_resolution) # Calculate layout aspect ratio for verification if position_data: x_rel, y_rel, w_rel, h_rel = position_data layout_aspect = w_rel / h_rel if h_rel > 0 else 1.0 image_aspect = width / height aspect_error = abs(image_aspect - layout_aspect) / layout_aspect * 100 print(f"Image {image_index + 1}/{num_images}: Layout aspect={layout_aspect:.4f}, Image aspect={image_aspect:.4f}, Error={aspect_error:.2f}%") # Set up the generator for reproducibility generator = torch.Generator(device="cuda").manual_seed(seed) print(f"Original prompt: '{prompt}'") print(f"Style preset: '{style_preset}'") print(f"Auto-selected size based on layout: {width}x{height}") # Apply style preset first styled_prompt, style_negative_prompt = apply_style_preset(prompt, style_preset, custom_style_text) # Add dialogue to the prompt if present if dialogue and dialogue.strip(): styled_prompt = f"{styled_prompt}. {dialogue.strip()}" # Use style negative prompt if available, otherwise default negative_prompt = style_negative_prompt if style_negative_prompt else " " print(f"Final Prompt: '{styled_prompt}'") print(f"Negative Prompt: '{negative_prompt}'") print(f"Seed: {seed}, Size: {width}x{height}, Steps: {num_inference_steps}, True CFG Scale: {guidance_scale}") # Generate the image image = pipe( prompt=styled_prompt, negative_prompt=negative_prompt, width=width, height=height, num_inference_steps=num_inference_steps, generator=generator, true_cfg_scale=guidance_scale, ).images[0] # Convert to grayscale if using manga_no_color style if style_preset == "manga_no_color": image = image.convert('L').convert('RGB') return image, seed # Keep the old infer function for backward compatibility (simplified) infer = infer_single_auto # --- Examples and UI Layout --- examples = [ "A capybara wearing a suit holding a sign that reads Hello World", ] css = """ #col-container { margin: 0 auto; max-width: 1024px; } #logo-title { text-align: center; } #logo-title img { width: 400px; } """ with gr.Blocks(css=css) as demo: gr.set_static_paths(paths=["logo.png"]) # Session state session_state = gr.State(value={"session_id": str(uuid.uuid4()), "page_count": 0}) with gr.Column(elem_id="col-container"): # Row 1: Logo, Inputs, Buttons with gr.Row(): # Column 1: Logo with gr.Column(scale=0, min_width=140): gr.HTML("""
AI Comic Factory Logo
""") # Column 2: Inputs with gr.Column(scale=1): # Sub-row 1: Prompt prompt = gr.Text( label="Prompt", show_label=False, placeholder="Describe the current page", container=False, ) # Sub-row 2: Style, number of images, layout with gr.Row(): # Create dropdown choices from loaded presets style_choices = [(preset["label"], key) for key, preset in STYLE_PRESETS.items()] style_preset = gr.Dropdown( label="Style Preset", choices=style_choices, value="no_style", interactive=True, scale=1 ) # Number of images slider num_images_slider = gr.Slider( label="Nb of panels", minimum=1, # we can support 8, but we will need better layouts # also starting from 8 I notice some latency for story generation maximum=7, step=1, value=6, scale=1 ) # Page layout dropdown - initialize with correct choices for default value (6 images) default_num_images = 6 default_layout_choices = get_layout_choices(default_num_images) layout_dropdown = gr.Dropdown( label="Page Layout", choices=default_layout_choices, value=default_layout_choices[0][1] if default_layout_choices else "default", interactive=True, scale=1 ) # Column 3: Buttons with gr.Column(scale=0,min_width=200): run_button = gr.Button("Generate page 1", variant="primary") reset_button = gr.Button("Reset", variant="secondary") # Row 2: Settings column and PDF preview with gr.Row(): # Left column - Additional settings with gr.Column(scale=0): custom_style_text = gr.Textbox( label="Custom Style Text", placeholder="Enter custom style (e.g., 'oil painting')", visible=False, lines=1 ) # Advanced settings accordion with gr.Accordion("Advanced Settings", open=False): guidance_scale = gr.Slider( label="Guidance scale (True CFG Scale)", minimum=1.0, maximum=5.0, step=0.1, value=1.0, ) num_inference_steps = gr.Slider( label="Number of inference steps", minimum=4, maximum=28, step=1, value=8, ) max_resolution = gr.Slider( label="Max Resolution", minimum=768, maximum=1280, step=128, value=1024, info="Maximum dimension for generated images (higher = better quality but slower)" ) gr.Markdown("""**Note:** Your images and PDF are saved for up to 24 hours. You can continue adding pages (up to 24) by clicking the generate button.""") # Right column - PDF Preview with gr.Column(scale=2): pdf_preview = PDF( label="PDF Preview", show_label=True, height=900, elem_id="pdf-preview", enable_zoom=True, min_zoom=1.0, max_zoom=3.0 ) # Add interaction to show/hide custom style text field def toggle_custom_style(style_value): return gr.update(visible=(style_value == "no_style")) style_preset.change( fn=toggle_custom_style, inputs=[style_preset], outputs=[custom_style_text] ) # Update layout dropdown when number of images changes def update_layout_choices(num_images): choices = get_layout_choices(int(num_images)) return gr.update(choices=choices, value=choices[0][1] if choices else "default") num_images_slider.change( fn=update_layout_choices, inputs=[num_images_slider], outputs=[layout_dropdown] ) # Define the main generation event generation_event = gr.on( triggers=[run_button.click, prompt.submit], fn=infer_page, inputs=[ prompt, guidance_scale, num_inference_steps, style_preset, custom_style_text, num_images_slider, layout_dropdown, max_resolution, session_state, ], outputs=[session_state, pdf_preview, run_button], ) # Reset button functionality def reset_session(): new_state = {"session_id": str(uuid.uuid4()), "page_count": 0} return new_state, None, None, "Generate page 1" # Connect the reset button reset_button.click( fn=reset_session, inputs=[], outputs=[session_state, pdf_preview, run_button] ) if __name__ == "__main__": demo.launch(mcp_server=True)