יסודות הבינה מלאכותית 12- RB34: יצירת סרטונים בעזרת בינה מלאכותית – מזה reinforcement learning – PPO הכי פשוט שאפשר
חדש בשכונה חדשות A.I
איך ניורולינק עובד – מחשב מוח ממשק
ניורולינק התקדמות
חלק א – מזה reinforcement learning – PPO הכי פשוט שאפשר
קוד הבינה מלאכותית
|
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 |
import os import time import gymnasium as gym import numpy as np from gymnasium import spaces from stable_baselines3 import PPO from stable_baselines3.common.callbacks import BaseCallback from stable_baselines3.common.env_checker import check_env # -------------------------------------------------- # SETTINGS # -------------------------------------------------- GRID_SIZE = 5 AGENT_ROW = 4 # Internal index 4 = displayed row 5 TRAINING_STEPS = 50_000 TEST_EPISODES = 1_000 MODEL_NAME = "falling_object_ppo" FALL_DELAY = 0.20 # Pause between falling rows FINAL_DELAY = 1.0 # Pause on reward/punishment ANIMATED_EPISODES = TRAINING_STEPS # All steps are displayed # -------------------------------------------------- # 5x5 ENVIRONMENT # -------------------------------------------------- class FallingObjectEnv(gym.Env): metadata = {"render_modes": ["human"]} def __init__(self): super().__init__() # Actions: # 0 = left # 1 = stay # 2 = right self.action_space = spaces.Discrete(3) # PPO receives: # 15 values = three visible rows × five columns # 1 value = agent column self.observation_space = spaces.Box( low=0.0, high=1.0, shape=(16,), dtype=np.float32, ) self.agent_column = 2 # Start at -1 so the first step moves the object to row 1. self.object_row = -1 self.object_column = 0 self.episode_number = 0 self.total_avoided = 0 self.total_collisions = 0 self.last_action = "NONE" self.last_reward = 0.0 self.last_result = "RUNNING" def reset(self, seed=None, options=None): super().reset(seed=seed) # Agent starts in the center. self.agent_column = 2 # Object begins above the matrix. # The first step moves it into displayed row 1. self.object_row = -1 self.object_column = int( self.np_random.integers( low=0, high=GRID_SIZE, ) ) self.last_action = "NONE" self.last_reward = 0.0 self.last_result = "RUNNING" return self.get_observation(), self.get_info() def step(self, action): action = int(action) action_names = { 0: "LEFT", 1: "STAY", 2: "RIGHT", } self.last_action = action_names[action] old_agent_column = self.agent_column # Move the agent. if action == 0: self.agent_column -= 1 elif action == 2: self.agent_column += 1 # Keep agent inside columns 1-5. self.agent_column = int( np.clip( self.agent_column, 0, GRID_SIZE - 1, ) ) # Small cost only when the agent really moves. if self.agent_column != old_agent_column: reward = -0.01 else: reward = 0.0 # Move object down by one row. self.object_row += 1 terminated = False truncated = False self.last_result = "RUNNING" # Object reached displayed row 5. if self.object_row == AGENT_ROW: terminated = True self.episode_number += 1 if self.object_column == self.agent_column: reward = -10.0 self.last_result = "COLLISION" self.total_collisions += 1 else: reward = 10.0 self.last_result = "AVOIDED" self.total_avoided += 1 self.last_reward = float(reward) observation = self.get_observation() info = self.get_info() return ( observation, float(reward), terminated, truncated, info, ) def get_observation(self): """ PPO sees only the three rows above the agent: Displayed row 2 = internal row 1 Displayed row 3 = internal row 2 Displayed row 4 = internal row 3 Displayed row 1 is hidden. """ visible_matrix = np.zeros( shape=(3, GRID_SIZE), dtype=np.float32, ) if 1 <= self.object_row <= 3: visible_row = self.object_row - 1 visible_matrix[ visible_row, self.object_column, ] = 1.0 normalized_agent_column = np.array( [ self.agent_column / (GRID_SIZE - 1) ], dtype=np.float32, ) observation = np.concatenate( [ visible_matrix.flatten(), normalized_agent_column, ] ) return observation.astype(np.float32) def get_info(self): """ Information used by the terminal display. PPO does not receive this information as its observation. """ return { "agent_column": self.agent_column, "object_row": self.object_row, "object_column": self.object_column, "episode_number": self.episode_number, "total_avoided": self.total_avoided, "total_collisions": self.total_collisions, "action_name": self.last_action, "reward": self.last_reward, "result": self.last_result, } # -------------------------------------------------- # TERMINAL FUNCTIONS # -------------------------------------------------- def clear_terminal(): os.system("cls" if os.name == "nt" else "clear") def create_grid_lines(info): """ Build the 5x5 matrix from the step information. This is important because Stable-Baselines3 automatically resets an environment when an episode finishes. """ agent_column = int(info["agent_column"]) object_row = int(info["object_row"]) object_column = int(info["object_column"]) grid = [ ["." for _ in range(GRID_SIZE)] for _ in range(GRID_SIZE) ] # Draw the object. if 0 <= object_row < GRID_SIZE: grid[object_row][object_column] = "O" collision = ( object_row == AGENT_ROW and object_column == agent_column ) # Draw agent or collision. if collision: grid[AGENT_ROW][agent_column] = "X" else: grid[AGENT_ROW][agent_column] = "A" lines = ["+-----------+"] for row_index, row in enumerate(grid): row_text = " ".join(row) displayed_row = row_index + 1 if row_index == 0: description = f"row {displayed_row} hidden" elif row_index < AGENT_ROW: description = f"row {displayed_row} visible" else: description = "row 5 agent" lines.append( f"| {row_text} | {description}" ) lines.append("+-----------+") return lines def print_side_by_side(left_lines, right_lines): total_lines = max( len(left_lines), len(right_lines), ) for index in range(total_lines): if index < len(left_lines): left_text = left_lines[index] else: left_text = "" if index < len(right_lines): right_text = right_lines[index] else: right_text = "" print( left_text.ljust(32) + right_text ) # -------------------------------------------------- # PPO TERMINAL CALLBACK # -------------------------------------------------- class TerminalTrainingCallback(BaseCallback): def __init__(self): super().__init__() def _on_step(self): info = self.locals["infos"][0] avoided = int(info.get("total_avoided", 0)) collisions = int(info.get("total_collisions", 0)) completed_episodes = avoided + collisions result = info.get("result", "RUNNING") reward = float(info.get("reward", 0.0)) if completed_episodes > 0: success_rate = avoided / completed_episodes * 100.0 else: success_rate = 0.0 progress = min( self.num_timesteps / TRAINING_STEPS * 100.0, 100.0, ) if reward == 10.0: reward_text = "+10 REWARD" elif reward == -10.0: reward_text = "-10 PUNISHMENT" elif reward < 0: reward_text = "-0.01 MOVE COST" else: reward_text = "0" left_lines = create_grid_lines(info) right_lines = [ "PPO LEARNING", "============", f"Training step: {self.num_timesteps:,}", f"Progress: {progress:6.2f}%", f"Episode: {completed_episodes + 1:,}", f"Avoided: {avoided:,}", f"Collisions: {collisions:,}", f"Success: {success_rate:6.2f}%", "", f"Action: {info.get('action_name', '-')}", f"Result: {result}", f"Reward: {reward_text}", ] clear_terminal() print( "5x5 WORLD".ljust(32) + "TRAINING INFORMATION" ) print() print_side_by_side( left_lines, right_lines, ) print() print("O = falling object") print("A = agent") print("X = collision") if result in ("AVOIDED", "COLLISION"): time.sleep(FINAL_DELAY) else: time.sleep(FALL_DELAY) return True # -------------------------------------------------- # TEST TRAINED MODEL # -------------------------------------------------- def evaluate_model(model, episodes): environment = FallingObjectEnv() avoided = 0 collisions = 0 total_reward = 0.0 for _ in range(episodes): observation, info = environment.reset() terminated = False truncated = False episode_reward = 0.0 while not (terminated or truncated): action, _ = model.predict( observation, deterministic=True, ) ( observation, reward, terminated, truncated, info, ) = environment.step(action) episode_reward += reward if info["result"] == "AVOIDED": avoided += 1 else: collisions += 1 total_reward += episode_reward environment.close() success_rate = ( avoided / episodes * 100.0 ) average_reward = ( total_reward / episodes ) return ( avoided, collisions, success_rate, average_reward, ) # -------------------------------------------------- # MAIN PROGRAM # -------------------------------------------------- def main(): environment = FallingObjectEnv() print("Checking environment...") check_env( environment, warn=True, ) policy_settings = { "net_arch": { # Actor network "pi": [32, 32], # Critic network "vf": [32, 32], } } model = PPO( policy="MlpPolicy", env=environment, learning_rate=0.0003, n_steps=256, batch_size=64, n_epochs=10, gamma=0.99, gae_lambda=0.95, clip_range=0.2, ent_coef=0.01, policy_kwargs=policy_settings, verbose=0, seed=42, device="auto", ) callback = TerminalTrainingCallback() print("Starting PPO training...") model.learn( total_timesteps=TRAINING_STEPS, callback=callback, ) model.save(MODEL_NAME) environment.close() ( avoided, collisions, success_rate, average_reward, ) = evaluate_model( model=model, episodes=TEST_EPISODES, ) clear_terminal() print("TRAINING FINISHED") print("=================") print() print(f"Model saved: {MODEL_NAME}.zip") print() print("TEST SUMMARY") print("------------") print(f"Test episodes: {TEST_EPISODES}") print(f"Avoided: {avoided}") print(f"Collisions: {collisions}") print(f"Success rate: {success_rate:.2f}%") print(f"Average reward: {average_reward:.3f}") print() if success_rate >= 95.0: print( "RESULT: The PPO agent learned very well." ) elif success_rate >= 85.0: print( "RESULT: The PPO agent learned." ) elif success_rate > 80.0: print( "RESULT: The PPO agent learned partially." ) else: print( "RESULT: The PPO agent did not learn enough." ) if __name__ == "__main__": main() |
יצירת סרטונים בעזרת בינה מלאכותית
-
כמה עולה פר סרטון של 10 שניות ?
-
איך שומרים על דמות בין הסרטונים ?
-
איך עושים כל מיני אפקטים ?
-
מי מודלי ה A.I המובלים ?
-
מעטפת – יתרונות חסרונות
חלק ב : כמה עולה ?
ממוצע 3 דולר ל 10 שניות – בין 2 ל 5 דולר ל בין 8 עד 15 שניות תלוי בחברה מודל
אבל זה יש להכפיל בכמות הניסיונות עד לקבלת המודל הרצוי
כמה עולה – Gemini Omni Flash in Google Flow הערכה
| Google plan | Monthly price | Credits | 8-second videos | 10-second videos |
|---|---|---|---|---|
| Free | $0 | 50 daily | 2 daily | 1 daily |
| AI Plus | $4.99 | 200/month | 8 | 6 |
| AI Pro | $19.99 | 1,000/month | 40 | 33 |
| AI Ultra 5x | $99.99 | 10,000/month | 400 | 333 |
| AI Ultra 20x | $199.99 | 25,000/month | 1,000 | 833 |
מחירי KLING
| מסלול | מחיר מבצע | קרדיטים | סרטוני 8 שניות | סרטוני 10 שניות |
|---|---|---|---|---|
| Standard | $6.99 | 660 | 6 | 5 |
| Pro | $25.99 | 3,000 | 31 | 25 |
| Premier | $64.99 | 8,000 | 83 | 66 |
| Ultra | $127.99 | 26,000 | 270 | 216 |
מחירי Luma Dream Machine
| מסלול Web | מחיר חודשי | קרדיטים | סרטוני 10 שניות Ray 3.2 Draft | סרטוני 10 שניות 720p |
|---|---|---|---|---|
| Lite | $9.99 | 3,200 | 53 | 10 |
| Plus | $29.99 | 10,000 | 166 | 33 |
| Unlimited | $94.99 | 10,000 מהירים | 166 מהירים | 33 מהירים |
| Unlimited Relaxed | כלול | ללא הגבלה | ללא הגבלה, איטי | ללא הגבלה, איטי |
שמירה על אותה הדמות בין סרטונים
CHARACTER CONSISTENCY
השיטה הטובה ביותר – CHARACTER CONSISTENCY
- צור דף דמות קבוע הכולל פנים, חזית, פרופיל, גוף מלא והלבוש.
- שמור את הדמות כ־Character/Element במערכת.
- בכל סרטון השתמש באותו קובץ דמות כ־Reference.
- שמור את הפריים האחרון של סרטון א׳ והשתמש בו כפריים הראשון של סרטון ב׳.
- חזור בדיוק על תיאור השיער, הבגדים, הגיל והאביזרים.
- צור קטעים קצרים של 5–10 שניות וחבר אותם בעריכה.
זה דוגמא : לדמות REFERANCE
CREATE ELEMENT
עבור דמות ראשונה :
עבור דמות הראשונה
עבור הדמות השנייה
הפריים האחרון הופך לפריים הראשון של הסרטון הבא.
| פלטפורמה | כלי שמירת הדמות | רמת עקביות |
|---|---|---|
| Kling 3.0 | Elements, מספר תמונות ייחוס ו־Storyboard | מצוינת |
| OpenArt | Consistent Character ומודל אישי | מצוינת מאוד |
| Luma | Master Reference ו־Keyframes | טובה מאוד |
| Google Flow/Veo | Ingredients, תמונות ייחוס ופריים ראשון/אחרון | טובה |
| טקסט בלבד | תיאור הדמות בכל Prompt | חלשה |
חלק ג OPEN ART מעטפת של מודלים
בניית אפקטים שלב אחרי שלב
1.2 שמירת אחידות לוגו SKILL brand kit
