| 1 | # Copyright (C) 2025 AIDC-AI |
| 2 | # |
| 3 | # Licensed under the Apache License, Version 2.0 (the "License"); |
| 4 | # you may not use this file except in compliance with the License. |
| 5 | # You may obtain a copy of the License at |
| 6 | # http://www.apache.org/licenses/LICENSE-2.0 |
| 7 | # Unless required by applicable law or agreed to in writing, software |
| 8 | # distributed under the License is distributed on an "AS IS" BASIS, |
| 9 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 10 | # See the License for the specific language governing permissions and |
| 11 | # limitations under the License. |
| 12 | |
| 13 | """ |
| 14 | Content generation utility functions |
| 15 | |
| 16 | Pure/stateless functions for generating content using LLM. |
| 17 | These functions are reusable across different pipelines. |
| 18 | """ |
| 19 | |
| 20 | import json |
| 21 | import re |
| 22 | from typing import List, Optional, Literal |
| 23 | |
| 24 | from loguru import logger |
| 25 | |
| 26 | |
| 27 | async def generate_title( |
| 28 | llm_service, |
| 29 | content: str, |
| 30 | strategy: Literal["auto", "direct", "llm"] = "auto", |
| 31 | max_length: int = 15 |
| 32 | ) -> str: |
| 33 | """ |
| 34 | Generate title from content |
| 35 | |
| 36 | Args: |
| 37 | llm_service: LLM service instance |
| 38 | content: Source content (topic or script) |
| 39 | strategy: Generation strategy |
| 40 | - "auto": Auto-decide based on content length (default) |
| 41 | - "direct": Use content directly (truncated if needed) |
| 42 | - "llm": Always use LLM to generate title |
| 43 | max_length: Maximum title length (default: 15) |
| 44 | |
| 45 | Returns: |
| 46 | Generated title |
| 47 | """ |
| 48 | if strategy == "direct": |
| 49 | content = content.strip() |
| 50 | return content[:max_length] if len(content) > max_length else content |
| 51 | |
| 52 | if strategy == "auto": |
| 53 | if len(content.strip()) <= 15: |
| 54 | return content.strip() |
| 55 | # Fall through to LLM |
| 56 | |
| 57 | # Use LLM to generate title |
| 58 | from pixelle_video.prompts import build_title_generation_prompt |
| 59 | |
| 60 | # Pass max_length to prompt so LLM knows the character limit |
| 61 | prompt = build_title_generation_prompt(content, max_length=max_length) |
| 62 | response = await llm_service(prompt, temperature=0.7, max_tokens=2000) |
| 63 | |
| 64 | # Clean up response |
| 65 | title = response.strip() |
| 66 | |
| 67 | # Remove quotes if present |
| 68 | if title.startswith('"') and title.endswith('"'): |
| 69 | title = title[1:-1] |
| 70 | if title.startswith("'") and title.endswith("'"): |
| 71 | title = title[1:-1] |
| 72 | |
| 73 | # Remove trailing punctuation |
| 74 | title = title.rstrip('.,!?;:\'"') |
| 75 | |
| 76 | # Safety: if still over limit, truncate smartly |
| 77 | if len(title) > max_length: |
| 78 | # Try to truncate at word boundary |
| 79 | truncated = title[:max_length] |
| 80 | last_space = truncated.rfind(' ') |
| 81 | |
| 82 | # Only use word boundary if it's not too far back (at least 60% of max_length) |
| 83 | if last_space > max_length * 0.6: |
| 84 | title = truncated[:last_space] |
| 85 | else: |
| 86 | title = truncated |
| 87 | |
| 88 | # Remove any trailing punctuation after truncation |
| 89 | title = title.rstrip('.,!?;:\'"') |
| 90 | |
| 91 | logger.debug(f"Generated title: '{title}' (length: {len(title)})") |
| 92 | return title |
| 93 | |
| 94 | |
| 95 | async def generate_narrations_from_topic( |
| 96 | llm_service, |
| 97 | topic: str, |
| 98 | n_scenes: int = 5, |
| 99 | min_words: int = 5, |
| 100 | max_words: int = 20 |
| 101 | ) -> List[str]: |
| 102 | """ |
| 103 | Generate narrations from topic using LLM |
| 104 | |
| 105 | Args: |
| 106 | llm_service: LLM service instance |
| 107 | topic: Topic/theme to generate narrations from |
| 108 | n_scenes: Number of narrations to generate |
| 109 | min_words: Minimum narration length |
| 110 | max_words: Maximum narration length |
| 111 | |
| 112 | Returns: |
| 113 | List of narration texts |
| 114 | """ |
| 115 | from pixelle_video.prompts import build_topic_narration_prompt |
| 116 | |
| 117 | logger.info(f"Generating {n_scenes} narrations from topic: {topic}") |
| 118 | |
| 119 | prompt = build_topic_narration_prompt( |
| 120 | topic=topic, |
| 121 | n_storyboard=n_scenes, |
| 122 | min_words=min_words, |
| 123 | max_words=max_words |
| 124 | ) |
| 125 | |
| 126 | response = await llm_service( |
| 127 | prompt=prompt, |
| 128 | temperature=0.8, |
| 129 | max_tokens=2000 |
| 130 | ) |
| 131 | |
| 132 | logger.debug(f"LLM response: {response[:200]}...") |
| 133 | |
| 134 | # Parse JSON |
| 135 | result = _parse_json(response) |
| 136 | |
| 137 | if "narrations" not in result: |
| 138 | raise ValueError("Invalid response format: missing 'narrations' key") |
| 139 | |
| 140 | narrations = result["narrations"] |
| 141 | |
| 142 | # Validate count |
| 143 | if len(narrations) > n_scenes: |
| 144 | logger.warning(f"Got {len(narrations)} narrations, taking first {n_scenes}") |
| 145 | narrations = narrations[:n_scenes] |
| 146 | elif len(narrations) < n_scenes: |
| 147 | raise ValueError(f"Expected {n_scenes} narrations, got only {len(narrations)}") |
| 148 | |
| 149 | logger.info(f"Generated {len(narrations)} narrations successfully") |
| 150 | return narrations |
| 151 | |
| 152 | |
| 153 | async def generate_narrations_from_content( |
| 154 | llm_service, |
| 155 | content: str, |
| 156 | n_scenes: int = 5, |
| 157 | min_words: int = 5, |
| 158 | max_words: int = 20 |
| 159 | ) -> List[str]: |
| 160 | """ |
| 161 | Generate narrations from user-provided content using LLM |
| 162 | |
| 163 | Args: |
| 164 | llm_service: LLM service instance |
| 165 | content: User-provided content |
| 166 | n_scenes: Number of narrations to generate |
| 167 | min_words: Minimum narration length |
| 168 | max_words: Maximum narration length |
| 169 | |
| 170 | Returns: |
| 171 | List of narration texts |
| 172 | """ |
| 173 | from pixelle_video.prompts import build_content_narration_prompt |
| 174 | |
| 175 | logger.info(f"Generating {n_scenes} narrations from content ({len(content)} chars)") |
| 176 | |
| 177 | prompt = build_content_narration_prompt( |
| 178 | content=content, |
| 179 | n_storyboard=n_scenes, |
| 180 | min_words=min_words, |
| 181 | max_words=max_words |
| 182 | ) |
| 183 | |
| 184 | response = await llm_service( |
| 185 | prompt=prompt, |
| 186 | temperature=0.8, |
| 187 | max_tokens=2000 |
| 188 | ) |
| 189 | |
| 190 | # Parse JSON |
| 191 | result = _parse_json(response) |
| 192 | |
| 193 | if "narrations" not in result: |
| 194 | raise ValueError("Invalid response format: missing 'narrations' key") |
| 195 | |
| 196 | narrations = result["narrations"] |
| 197 | |
| 198 | # Validate count |
| 199 | if len(narrations) > n_scenes: |
| 200 | logger.warning(f"Got {len(narrations)} narrations, taking first {n_scenes}") |
| 201 | narrations = narrations[:n_scenes] |
| 202 | elif len(narrations) < n_scenes: |
| 203 | raise ValueError(f"Expected {n_scenes} narrations, got only {len(narrations)}") |
| 204 | |
| 205 | logger.info(f"Generated {len(narrations)} narrations successfully") |
| 206 | return narrations |
| 207 | |
| 208 | |
| 209 | async def split_narration_script( |
| 210 | script: str, |
| 211 | split_mode: Literal["paragraph", "line", "sentence"] = "paragraph", |
| 212 | ) -> List[str]: |
| 213 | """ |
| 214 | Split user-provided narration script into segments |
| 215 | |
| 216 | Args: |
| 217 | script: Fixed narration script |
| 218 | split_mode: Splitting strategy |
| 219 | - "paragraph": Split by double newline (\\n\\n), preserve single newlines within paragraphs |
| 220 | - "line": Split by single newline (\\n), each line is a segment |
| 221 | - "sentence": Split by sentence-ending punctuation (。.!?！？) |
| 222 | |
| 223 | Returns: |
| 224 | List of narration segments |
| 225 | """ |
| 226 | logger.info(f"Splitting script (mode={split_mode}, length={len(script)} chars)") |
| 227 | |
| 228 | narrations = [] |
| 229 | |
| 230 | if split_mode == "paragraph": |
| 231 | # Split by double newline (paragraph mode) |
| 232 | # Preserve single newlines within paragraphs |
| 233 | paragraphs = re.split(r'\n\s*\n', script) |
| 234 | for para in paragraphs: |
| 235 | # Only strip leading/trailing whitespace, preserve internal newlines |
| 236 | cleaned = para.strip() |
| 237 | if cleaned: |
| 238 | narrations.append(para) |
| 239 | logger.info(f"✅ Split script into {len(narrations)} segments (by paragraph)") |
| 240 | |
| 241 | elif split_mode == "line": |
| 242 | # Split by single newline (original behavior) |
| 243 | narrations = [line.strip() for line in script.split('\n') if line.strip()] |
| 244 | logger.info(f"✅ Split script into {len(narrations)} segments (by line)") |
| 245 | |
| 246 | elif split_mode == "sentence": |
| 247 | # Split by sentence-ending punctuation |
| 248 | # Supports Chinese (。！？) and English (.!?) |
| 249 | # Use regex to split while keeping sentences intact |
| 250 | cleaned = re.sub(r'\s+', ' ', script.strip()) |
| 251 | # Split on sentence-ending punctuation, keeping the punctuation with the sentence |
| 252 | sentences = re.split(r'(?<=[。.!?！？])\s*', cleaned) |
| 253 | narrations = [s.strip() for s in sentences if s.strip()] |
| 254 | logger.info(f"✅ Split script into {len(narrations)} segments (by sentence)") |
| 255 | |
| 256 | else: |
| 257 | # Fallback to line mode |
| 258 | logger.warning(f"Unknown split_mode '{split_mode}', falling back to 'line'") |
| 259 | narrations = [line.strip() for line in script.split('\n') if line.strip()] |
| 260 | |
| 261 | # Log statistics |
| 262 | if narrations: |
| 263 | lengths = [len(s) for s in narrations] |
| 264 | logger.info(f" Min: {min(lengths)} chars, Max: {max(lengths)} chars, Avg: {sum(lengths)//len(lengths)} chars") |
| 265 | |
| 266 | return narrations |
| 267 | |
| 268 | |
| 269 | async def generate_image_prompts( |
| 270 | llm_service, |
| 271 | narrations: List[str], |
| 272 | min_words: int = 30, |
| 273 | max_words: int = 60, |
| 274 | batch_size: int = 10, |
| 275 | max_retries: int = 3, |
| 276 | progress_callback: Optional[callable] = None |
| 277 | ) -> List[str]: |
| 278 | """ |
| 279 | Generate image prompts from narrations (with batching and retry) |
| 280 | |
| 281 | Args: |
| 282 | llm_service: LLM service instance |
| 283 | narrations: List of narrations |
| 284 | min_words: Min image prompt length |
| 285 | max_words: Max image prompt length |
| 286 | batch_size: Max narrations per batch (default: 10) |
| 287 | max_retries: Max retry attempts per batch (default: 3) |
| 288 | progress_callback: Optional callback(completed, total, message) for progress updates |
| 289 | |
| 290 | Returns: |
| 291 | List of image prompts (base prompts, without prefix applied) |
| 292 | """ |
| 293 | from pixelle_video.prompts import build_image_prompt_prompt |
| 294 | |
| 295 | logger.info(f"Generating image prompts for {len(narrations)} narrations (batch_size={batch_size})") |
| 296 | |
| 297 | # Split narrations into batches |
| 298 | batches = [narrations[i:i + batch_size] for i in range(0, len(narrations), batch_size)] |
| 299 | logger.info(f"Split into {len(batches)} batches") |
| 300 | |
| 301 | all_prompts = [] |
| 302 | |
| 303 | # Process each batch |
| 304 | for batch_idx, batch_narrations in enumerate(batches, 1): |
| 305 | logger.info(f"Processing batch {batch_idx}/{len(batches)} ({len(batch_narrations)} narrations)") |
| 306 | |
| 307 | # Retry logic for this batch |
| 308 | for attempt in range(1, max_retries + 1): |
| 309 | try: |
| 310 | # Generate prompts for this batch |
| 311 | prompt = build_image_prompt_prompt( |
| 312 | narrations=batch_narrations, |
| 313 | min_words=min_words, |
| 314 | max_words=max_words |
| 315 | ) |
| 316 | |
| 317 | response = await llm_service( |
| 318 | prompt=prompt, |
| 319 | temperature=0.7, |
| 320 | max_tokens=8192 |
| 321 | ) |
| 322 | |
| 323 | logger.debug(f"Batch {batch_idx} attempt {attempt}: LLM response length: {len(response)} chars") |
| 324 | |
| 325 | # Parse JSON |
| 326 | result = _parse_json(response) |
| 327 | |
| 328 | if "image_prompts" not in result: |
| 329 | raise KeyError("Invalid response format: missing 'image_prompts'") |
| 330 | |
| 331 | batch_prompts = result["image_prompts"] |
| 332 | |
| 333 | # Validate count |
| 334 | if len(batch_prompts) != len(batch_narrations): |
| 335 | error_msg = ( |
| 336 | f"Batch {batch_idx} prompt count mismatch (attempt {attempt}/{max_retries}):\n" |
| 337 | f" Expected: {len(batch_narrations)} prompts\n" |
| 338 | f" Got: {len(batch_prompts)} prompts" |
| 339 | ) |
| 340 | logger.warning(error_msg) |
| 341 | |
| 342 | if attempt < max_retries: |
| 343 | logger.info(f"Retrying batch {batch_idx}...") |
| 344 | continue |
| 345 | else: |
| 346 | raise ValueError(error_msg) |
| 347 | |
| 348 | # Success! |
| 349 | logger.info(f"✅ Batch {batch_idx} completed successfully ({len(batch_prompts)} prompts)") |
| 350 | all_prompts.extend(batch_prompts) |
| 351 | |
| 352 | # Report progress |
| 353 | if progress_callback: |
| 354 | progress_callback( |
| 355 | len(all_prompts), |
| 356 | len(narrations), |
| 357 | f"Batch {batch_idx}/{len(batches)} completed" |
| 358 | ) |
| 359 | |
| 360 | break |
| 361 | |
| 362 | except json.JSONDecodeError as e: |
| 363 | logger.error(f"Batch {batch_idx} JSON parse error (attempt {attempt}/{max_retries}): {e}") |
| 364 | if attempt >= max_retries: |
| 365 | raise |
| 366 | logger.info(f"Retrying batch {batch_idx}...") |
| 367 | |
| 368 | logger.info(f"✅ Generated {len(all_prompts)} image prompts") |
| 369 | return all_prompts |
| 370 | |
| 371 | |
| 372 | async def generate_video_prompts( |
| 373 | llm_service, |
| 374 | narrations: List[str], |
| 375 | min_words: int = 30, |
| 376 | max_words: int = 60, |
| 377 | batch_size: int = 10, |
| 378 | max_retries: int = 3, |
| 379 | progress_callback: Optional[callable] = None |
| 380 | ) -> List[str]: |
| 381 | """ |
| 382 | Generate video prompts from narrations (with batching and retry) |
| 383 | |
| 384 | Args: |
| 385 | llm_service: LLM service instance |
| 386 | narrations: List of narrations |
| 387 | min_words: Min video prompt length |
| 388 | max_words: Max video prompt length |
| 389 | batch_size: Max narrations per batch (default: 10) |
| 390 | max_retries: Max retry attempts per batch (default: 3) |
| 391 | progress_callback: Optional callback(completed, total, message) for progress updates |
| 392 | |
| 393 | Returns: |
| 394 | List of video prompts (base prompts, without prefix applied) |
| 395 | """ |
| 396 | from pixelle_video.prompts.video_generation import build_video_prompt_prompt |
| 397 | |
| 398 | logger.info(f"Generating video prompts for {len(narrations)} narrations (batch_size={batch_size})") |
| 399 | |
| 400 | # Split narrations into batches |
| 401 | batches = [narrations[i:i + batch_size] for i in range(0, len(narrations), batch_size)] |
| 402 | logger.info(f"Split into {len(batches)} batches") |
| 403 | |
| 404 | all_prompts = [] |
| 405 | |
| 406 | # Process each batch |
| 407 | for batch_idx, batch_narrations in enumerate(batches, 1): |
| 408 | logger.info(f"Processing batch {batch_idx}/{len(batches)} ({len(batch_narrations)} narrations)") |
| 409 | |
| 410 | # Retry logic for this batch |
| 411 | for attempt in range(1, max_retries + 1): |
| 412 | try: |
| 413 | # Generate prompts for this batch |
| 414 | prompt = build_video_prompt_prompt( |
| 415 | narrations=batch_narrations, |
| 416 | min_words=min_words, |
| 417 | max_words=max_words |
| 418 | ) |
| 419 | |
| 420 | response = await llm_service( |
| 421 | prompt=prompt, |
| 422 | temperature=0.7, |
| 423 | max_tokens=8192 |
| 424 | ) |
| 425 | |
| 426 | logger.debug(f"Batch {batch_idx} attempt {attempt}: LLM response length: {len(response)} chars") |
| 427 | |
| 428 | # Parse JSON |
| 429 | result = _parse_json(response) |
| 430 | |
| 431 | if "video_prompts" not in result: |
| 432 | raise KeyError("Invalid response format: missing 'video_prompts'") |
| 433 | |
| 434 | batch_prompts = result["video_prompts"] |
| 435 | |
| 436 | # Validate batch result |
| 437 | if len(batch_prompts) != len(batch_narrations): |
| 438 | raise ValueError( |
| 439 | f"Prompt count mismatch: expected {len(batch_narrations)}, got {len(batch_prompts)}" |
| 440 | ) |
| 441 | |
| 442 | # Success - add to all_prompts |
| 443 | all_prompts.extend(batch_prompts) |
| 444 | logger.info(f"✓ Batch {batch_idx} completed: {len(batch_prompts)} video prompts") |
| 445 | |
| 446 | # Report progress |
| 447 | if progress_callback: |
| 448 | completed = len(all_prompts) |
| 449 | total = len(narrations) |
| 450 | progress_callback(completed, total, f"Batch {batch_idx}/{len(batches)} completed") |
| 451 | |
| 452 | break # Success, move to next batch |
| 453 | |
| 454 | except Exception as e: |
| 455 | logger.warning(f"✗ Batch {batch_idx} attempt {attempt} failed: {e}") |
| 456 | if attempt >= max_retries: |
| 457 | raise |
| 458 | logger.info(f"Retrying batch {batch_idx}...") |
| 459 | |
| 460 | logger.info(f"✅ Generated {len(all_prompts)} video prompts") |
| 461 | return all_prompts |
| 462 | |
| 463 | |
| 464 | def _parse_json(text: str) -> dict: |
| 465 | """ |
| 466 | Parse JSON from text, with fallback to extract JSON from markdown code blocks |
| 467 | |
| 468 | Args: |
| 469 | text: Text containing JSON |
| 470 | |
| 471 | Returns: |
| 472 | Parsed JSON dict |
| 473 | |
| 474 | Raises: |
| 475 | json.JSONDecodeError: If no valid JSON found |
| 476 | """ |
| 477 | # Try direct parsing first |
| 478 | try: |
| 479 | return json.loads(text) |
| 480 | except json.JSONDecodeError: |
| 481 | pass |
| 482 | |
| 483 | # Try to extract JSON from markdown code block |
| 484 | json_pattern = r'```(?:json)?\s*([\s\S]+?)\s*```' |
| 485 | match = re.search(json_pattern, text, re.DOTALL) |
| 486 | if match: |
| 487 | try: |
| 488 | return json.loads(match.group(1)) |
| 489 | except json.JSONDecodeError: |
| 490 | pass |
| 491 | |
| 492 | # Try to find any JSON object in the text |
| 493 | json_pattern = r'\{[^{}]*(?:"narrations"|"image_prompts")\s*:\s*\[[^\]]*\][^{}]*\}' |
| 494 | match = re.search(json_pattern, text, re.DOTALL) |
| 495 | if match: |
| 496 | try: |
| 497 | return json.loads(match.group(0)) |
| 498 | except json.JSONDecodeError: |
| 499 | pass |
| 500 | |
| 501 | # If all fails, raise error |
| 502 | raise json.JSONDecodeError("No valid JSON found", text, 0) |
| 503 |