Spaces:
Running on Zero
Running on Zero
File size: 101,168 Bytes
6e4b62e | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 1493 1494 1495 1496 1497 1498 1499 1500 1501 1502 1503 1504 1505 1506 1507 1508 1509 1510 1511 1512 1513 1514 1515 1516 1517 1518 1519 1520 1521 1522 1523 1524 1525 1526 1527 1528 1529 1530 1531 1532 1533 1534 1535 1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 1550 1551 1552 1553 1554 1555 1556 1557 1558 1559 1560 1561 1562 1563 1564 1565 1566 1567 1568 1569 1570 1571 1572 1573 1574 1575 1576 1577 1578 1579 1580 1581 1582 1583 1584 1585 1586 1587 1588 1589 1590 1591 1592 1593 1594 1595 1596 1597 1598 1599 1600 1601 1602 1603 1604 1605 1606 1607 1608 1609 1610 1611 1612 1613 1614 1615 1616 1617 1618 1619 1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632 1633 1634 1635 1636 1637 1638 1639 1640 1641 1642 1643 1644 1645 1646 1647 1648 1649 1650 1651 1652 1653 1654 1655 1656 1657 1658 1659 1660 1661 1662 1663 1664 1665 1666 1667 1668 1669 1670 1671 1672 1673 1674 1675 1676 1677 1678 1679 1680 1681 1682 1683 1684 1685 1686 1687 1688 1689 1690 1691 1692 1693 1694 1695 1696 1697 1698 1699 1700 1701 1702 1703 1704 1705 1706 1707 1708 1709 1710 1711 1712 1713 1714 1715 1716 1717 1718 1719 1720 1721 1722 1723 1724 1725 1726 1727 1728 1729 1730 1731 1732 1733 1734 1735 1736 1737 1738 1739 1740 1741 1742 1743 1744 1745 1746 1747 1748 1749 1750 1751 1752 1753 1754 1755 1756 1757 1758 1759 1760 1761 1762 1763 1764 1765 1766 1767 1768 1769 1770 1771 1772 1773 1774 1775 1776 1777 1778 1779 1780 1781 1782 1783 1784 1785 1786 1787 1788 1789 1790 1791 1792 1793 1794 1795 1796 1797 1798 1799 1800 1801 1802 1803 1804 1805 1806 1807 1808 1809 1810 1811 1812 1813 1814 1815 1816 1817 1818 1819 1820 1821 1822 1823 1824 1825 1826 1827 1828 1829 1830 1831 1832 1833 1834 1835 1836 1837 1838 1839 1840 1841 1842 1843 1844 1845 1846 1847 1848 1849 1850 1851 1852 1853 1854 1855 1856 1857 1858 1859 1860 1861 1862 1863 1864 1865 1866 1867 1868 1869 1870 1871 1872 1873 1874 1875 1876 1877 1878 1879 1880 1881 1882 1883 1884 1885 1886 1887 1888 1889 1890 1891 1892 1893 1894 1895 1896 1897 1898 1899 1900 1901 1902 1903 1904 1905 1906 1907 1908 1909 1910 1911 1912 1913 1914 1915 1916 1917 1918 1919 1920 1921 1922 1923 1924 1925 1926 1927 1928 1929 1930 1931 1932 1933 1934 1935 1936 1937 1938 1939 1940 1941 1942 1943 1944 1945 1946 | #!/usr/bin/env python3
"""
回环:两栏输出 [输入轨迹 | 生成视频]。context 一律放在右侧(suffix),I2V 首帧也放在最右作为 clean 去噪 target。
- 第一帧:从训练数据随机采样一帧作为首 chunk 的 context(I2V,1 帧放右侧)
- 之后每 chunk 的第一帧(context)= 上一 chunk 的 last n 帧,同样放在该段生成视频的右侧
- 右栏视频 = 每段 [生成 81 帧 | context 帧],context 始终在右
默认回环两场景(每 chunk 45°):1) 1左1右 2chunk 2) 2左2右 4chunk。--run_legacy_loop 时跑旧三组(roundtrip/onedir/replay)。多卡按 rank 划分样本。
多 chunk 时 RT(相对位姿)与训练一致:
- 训练:ref_rt = rt_list[0],target_actions = convert_rt_to_relative(rt_list, ref_rt),即每段内 action 相对本段首帧。
- 推理:每个 chunk 的 action 也是相对该 chunk 的首帧。chunk1 注入 0°→45° 表示本段内转 45°;
chunk2 的首帧 = chunk1 的末帧(已 45°),chunk2 再注入 0°→45° 表示在 chunk2 局部再转 45°,世界系共 90°。
因此是 45+45,不是第二个 chunk 直接 90°(直接 90° 表示相对 chunk2 首帧转 90°,世界系会变成 45+90=135°)。
注意:条件以 action + context 为主;prompt 使用 GT 帧对应视频的文案以补充场景描述,利于生成清晰度。
"""
import os
import sys
import json
import argparse
import random
import math
import io
from datetime import datetime
import torch
import numpy as np
from PIL import Image
_script_dir = os.path.dirname(os.path.abspath(__file__))
_repo_root = os.path.dirname(os.path.dirname(_script_dir))
_exp1_4_2_dir = os.path.join(_repo_root, "ab_study", "exp1_4_2_context_suffix_cam_rt_relative")
if _repo_root not in sys.path:
sys.path.insert(0, _repo_root)
if _exp1_4_2_dir not in sys.path:
sys.path.insert(0, _exp1_4_2_dir)
if _script_dir not in sys.path:
sys.path.insert(0, _script_dir)
import loop_utils as irc
from diffsynth import save_video
from src.model_training.fov_retrieval import compute_rotation_list
from src.model_training.multichunk_sample_utils import (
context_frames_for_next_chunk,
replay_context_from_generated_frames,
)
def encode_context_frames_per_frame(pipe, pil_list, device, dtype=torch.bfloat16):
"""与训练 context_per_frame_vae 一致:每帧单独 VAE 编码再在时间维 concat,不做时序降采样。ctx=K → K 个 latent tokens(5/20 等均一帧一过 VAE)。"""
if not pil_list:
return None
encoded = []
for pil in pil_list:
frame_video = pipe.preprocess_video([pil]).to(device=device)
frame_sq = frame_video.squeeze(0) if frame_video.dim() == 5 else frame_video
if frame_sq.dim() == 3:
frame_sq = frame_sq.unsqueeze(0)
lat_one = pipe.vae.encode([frame_sq], device=pipe.device, tiled=False, tile_size=None, tile_stride=None)
encoded.append(lat_one)
context_latents = torch.cat(encoded, dim=2).to(dtype=dtype, device=device)
return context_latents
def _yaw_deg_from_rt(rt_list):
"""从 12 维 RT [t_x,t_y,t_z, R11,R12,...,R33] 提取 yaw(度)。R_z 时 yaw = atan2(R21, R11)."""
if not rt_list or len(rt_list) < 12:
return 0.0
R11, R21 = float(rt_list[3]), float(rt_list[6])
return math.degrees(math.atan2(R21, R11))
def build_action_ccw_cw(deg: float, chunk_frames: int = 81):
"""R_z(yaw),相对本段首帧。CCW=0→+deg,CW=0→-deg。匀速。"""
denom = max(1, chunk_frames - 1)
actions_ccw = {}
for i in range(chunk_frames):
yaw = (i / denom) * deg
actions_ccw[str(i)] = compute_rotation_list([0.0, 0.0, 0.0, yaw])
actions_cw = {}
for i in range(chunk_frames):
yaw = -(i / denom) * deg
actions_cw[str(i)] = compute_rotation_list([0.0, 0.0, 0.0, yaw])
return actions_ccw, actions_cw
def build_action_cw_then_ccw(deg: float, chunk_frames: int = 81):
"""匀速:chunk1 顺时针 0→-deg,chunk2 逆时针 0→+deg(本段局部)。折返:先 CW 30° 再 CCW 30° 回起点。"""
denom = max(1, chunk_frames - 1)
ch1 = {}
for i in range(chunk_frames):
ch1[str(i)] = compute_rotation_list([0.0, 0.0, 0.0, -(i / denom) * deg])
ch2 = {}
for i in range(chunk_frames):
ch2[str(i)] = compute_rotation_list([0.0, 0.0, 0.0, (i / denom) * deg])
return ch1, ch2
def build_gt_trajectory_actions(dataset_base, video_name, start_frame, chunk_frames, json_file=None):
"""与 Replay 完全一致:从数据集 json 取同段 GT pose,转成 relative RT 作为 action。若帧不足则返回 None。"""
if json_file is None:
json_file = os.path.join(dataset_base, "jsons", f"{video_name}.json")
if not os.path.isfile(json_file):
return None
try:
rt_list = [irc.load_pose_rt(json_file, start_frame + i) for i in range(chunk_frames)]
if not rt_list or any(r is None or len(r) < 12 for r in rt_list):
return None
ref_rt = rt_list[0]
rel_actions = {str(i): irc.get_relative_rt(rt_list[i], ref_rt) for i in range(chunk_frames)}
return rel_actions
except Exception:
return None
def build_action_from_gt_yaw_profile(dataset_base, video_name, start_frame, chunk_frames, target_total_deg, clockwise=False, json_file=None):
"""参考训练集方式:用同段 GT 轨迹的 yaw 曲线形状(时间分布)缩放为目标角度,使旋转节奏与训练集一致,而非简单线性。
若该段无 GT 或总转角过小则返回 None,调用方回退线性合成。"""
if json_file is None:
json_file = os.path.join(dataset_base, "jsons", f"{video_name}.json")
if not os.path.isfile(json_file):
return None
try:
rt_list = [irc.load_pose_rt(json_file, start_frame + i) for i in range(chunk_frames)]
if not rt_list or any(r is None or len(r) < 12 for r in rt_list):
return None
ref_rt = rt_list[0]
# GT 相对首帧的 yaw 曲线(与训练一致)
yaw_deg = [_yaw_deg_from_rt(irc.get_relative_rt(rt_list[i], ref_rt)) for i in range(chunk_frames)]
total_gt_deg = yaw_deg[-1] - yaw_deg[0] if len(yaw_deg) > 1 else 0.0
if abs(total_gt_deg) < 1e-4:
return None # 几乎无旋转,用线性更稳
# 归一化曲线 t[i] in [0,1],再缩放到 target_total_deg,保持 GT 的时间节奏
scale = target_total_deg / total_gt_deg
sign = -1 if clockwise else 1
actions = {}
for i in range(chunk_frames):
t = (yaw_deg[i] - yaw_deg[0]) / total_gt_deg # 0 -> 1
yaw = sign * t * target_total_deg
actions[str(i)] = compute_rotation_list([0.0, 0.0, 0.0, yaw])
return actions
except Exception:
return None
def build_action_chunk(deg: float, clockwise: bool, chunk_frames: int = 81):
"""单 chunk:0→deg(逆时针)或 0→-deg(顺时针)。相对本 chunk 首帧,与训练一致。
多 chunk 时:chunk2 的首帧=chunk1 的末帧,所以 chunk2 注入 0→deg 表示在 chunk2 局部再转 deg;
世界系转角 = chunk1 末帧 yaw + deg。例如 chunk1=45°,chunk2 再 45° 则注入 0→45°(不是 0→90°)。
返回 (actions_dict, yaw_list_deg)."""
denom = max(1, chunk_frames - 1)
actions = {}
yaw_list = []
for i in range(chunk_frames):
yaw = (i / denom) * (-deg if clockwise else deg)
yaw_list.append(yaw)
actions[str(i)] = compute_rotation_list([0.0, 0.0, 0.0, yaw])
return actions, yaw_list
def build_action_translation_only(direction: str, translation_delta: float, chunk_frames: int = 81):
"""纯平移单 chunk:R=I,沿单轴线性位移。与训练一致:fov_retrieval.pose_to_rt(..., constrain_to_xy=True) 仅用 XY,tz=0。
direction in ('forward','backward','left','right'):forward=+Y, backward=-Y, left=-X, right=+X(XY 平面,无 Z)。
RT 格式:前 3 维 [tx,ty,tz],后 9 维 3x3 行优先单位阵;相对本段首帧(与 convert_rt_to_relative(ref=首帧) 一致)。"""
# 与训练一致:2D 平面位移,z 恒为 0(见 fov_retrieval.pose_to_rt constrain_to_xy=True)
identity_rot = [1.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 1.0]
denom = max(1, chunk_frames - 1)
actions = {}
yaw_list = [0.0] * chunk_frames
for i in range(chunk_frames):
t = (i / denom) * translation_delta
if direction == "forward":
tx, ty, tz = 0.0, t, 0.0
elif direction == "backward":
tx, ty, tz = 0.0, -t, 0.0
elif direction == "left":
tx, ty, tz = -t, 0.0, 0.0
elif direction == "right":
tx, ty, tz = t, 0.0, 0.0
else:
tx, ty, tz = 0.0, 0.0, 0.0
actions[str(i)] = [tx, ty, tz] + identity_rot
return actions, yaw_list
def _draw_trajectory_pil(yaw_list, plot_width, plot_height, frame_index):
"""PIL-only fallback: draw trajectory 0..frame_index. Returns one PIL Image."""
pad = 40
w, h = plot_width - 2 * pad, plot_height - 2 * pad
if w <= 0 or h <= 0:
return Image.new("RGB", (plot_width, plot_height), (50, 50, 50))
img = Image.new("RGB", (plot_width, plot_height), (45, 45, 48))
end = min(frame_index + 1, len(yaw_list))
if end == 0:
return img
ys = [float(yaw_list[i]) for i in range(end)]
y_min, y_max = min(ys), max(ys)
if y_max <= y_min:
y_min, y_max = y_min - 10.0, y_max + 10.0
from PIL import ImageDraw
draw = ImageDraw.Draw(img)
n_x = max(1, len(yaw_list))
pts = []
for i in range(end):
x = pad + int((i / max(1, n_x - 1)) * w) if n_x > 1 else pad
yy = (ys[i] - y_min) / max(1e-6, y_max - y_min)
y = pad + int((1 - yy) * h)
pts.append((x, y))
if len(pts) >= 2:
draw.line(pts, fill=(0, 200, 255), width=2)
if pts:
draw.ellipse([pts[-1][0] - 4, pts[-1][1] - 4, pts[-1][0] + 4, pts[-1][1] + 4], fill=(255, 220, 0), outline=(255, 255, 255))
return img
def draw_trajectory_frames(yaw_history_deg, plot_width, plot_height, total_frames=None):
"""
yaw_history_deg: list of cumulative yaw (one per frame).
Returns list of PIL images (one per frame), each showing trajectory from 0 to current frame.
Uses matplotlib; if output is too dark, falls back to PIL-only drawing.
"""
yaw_list = [float(y) for y in (yaw_history_deg or [])]
n = len(yaw_list)
if total_frames is not None and total_frames > n:
n = total_frames
if n == 0:
return [Image.new("RGB", (plot_width, plot_height), (45, 45, 48))]
use_pil_fallback = False
try:
import matplotlib
matplotlib.use("Agg")
import matplotlib.pyplot as plt
except ImportError:
use_pil_fallback = True
if not use_pil_fallback:
y_min = min(yaw_list) if yaw_list else 0.0
y_max = max(yaw_list) if yaw_list else 0.0
if y_max <= y_min:
y_min, y_max = y_min - 15.0, y_max + 15.0
margin = max(5, (y_max - y_min) * 0.15)
y_lo, y_hi = y_min - margin, y_max + margin
out = []
for t in range(n):
fig, ax = plt.subplots(1, 1, figsize=(plot_width / 100.0, plot_height / 100.0), dpi=100)
fig.patch.set_facecolor("#252525")
ax.set_facecolor("#252525")
ax.set_xlim(0, max(n, 1))
ax.set_ylim(y_lo, y_hi)
ax.tick_params(colors="0.9", labelsize=7)
for spine in ax.spines.values():
spine.set_color("0.6")
end = min(t + 1, len(yaw_list))
if end > 0:
x = list(range(end))
y = [yaw_list[i] for i in range(end)]
ax.plot(x, y, color="cyan", linewidth=2.0, zorder=2)
cur_y = yaw_list[min(t, len(yaw_list) - 1)] if t < len(yaw_list) else (yaw_list[-1] if yaw_list else 0)
ax.scatter([t], [cur_y], color="yellow", s=35, zorder=3)
ax.set_xlabel("Frame", color="0.9", fontsize=8)
ax.set_ylabel("Yaw (deg)", color="0.9", fontsize=8)
ax.set_title("Trajectory (yaw)", color="0.95", fontsize=9)
fig.tight_layout(pad=0.5)
buf = io.BytesIO()
plt.savefig(buf, format="png", dpi=100, facecolor="#252525", edgecolor="none", bbox_inches="tight", pad_inches=0.2)
plt.close(fig)
buf.seek(0)
img = Image.open(buf).convert("RGB")
if img.size != (plot_width, plot_height):
img = img.resize((plot_width, plot_height), Image.Resampling.LANCZOS)
# If image is too dark (matplotlib failed in headless?), use PIL fallback for rest
arr = np.array(img)
if arr.mean() < 50:
use_pil_fallback = True
out = [_draw_trajectory_pil(yaw_list, plot_width, plot_height, t) for t in range(n)]
break
out.append(img)
if not use_pil_fallback:
return out
if use_pil_fallback:
return [_draw_trajectory_pil(yaw_list, plot_width, plot_height, t) for t in range(n)]
return out
def composite_two_panel(trajectory_frames, right_panel_frames, w_traj=320, h_traj=352, w_gen=640, h_gen=352):
"""左:输入轨迹可视化;右:生成视频(每段后接 context,context 在右)。"""
n = len(right_panel_frames)
if not trajectory_frames or len(trajectory_frames) != n:
traj_frames = draw_trajectory_frames([], w_traj, h_traj, total_frames=n)
else:
traj_frames = trajectory_frames
out = []
for i in range(n):
traj = traj_frames[i] if i < len(traj_frames) else traj_frames[-1]
if traj.size != (w_traj, h_traj):
traj = traj.resize((w_traj, h_traj))
gen = _frame_to_pil(right_panel_frames[i], w_gen, h_gen)
canvas_w = w_traj + w_gen
canvas_h = max(h_traj, h_gen)
canvas = Image.new("RGB", (canvas_w, canvas_h), (40, 40, 40))
canvas.paste(traj, (0, 0))
canvas.paste(gen, (w_traj, 0))
out.append(canvas)
return out
def load_sample_first_frame(dataset_base, video_name, start_frame, w, h):
"""加载当前样本的首帧(与 Replay 一致:context = 同条轨迹同视频的首帧),避免 loop 用随机帧导致 prompt 与 context 场景不一致而糊。返回 PIL 或 None。"""
if not video_name or start_frame is None:
return None
for fmt in (f"{int(start_frame):04d}.png", f"{int(start_frame)}.png"):
img_p = os.path.join(dataset_base, "frames", str(video_name), fmt)
if os.path.isfile(img_p):
return Image.open(img_p).convert("RGB").resize((w, h))
return None
def load_gt_frames_at_indices(dataset_base, video_name, frame_indices, w, h):
"""从数据集加载指定帧序号的 GT 图像,用于 chunk 间衔接的 context(与训练一致:context=真实帧,避免用生成帧再编码导致糊)。
frame_indices: 从左到右使用(如 [80, 79] 表示最后一帧、倒数第二帧)。返回 list[PIL] 若全部存在,否则 None。"""
if not video_name or not frame_indices:
return None
out = []
base = os.path.join(dataset_base, "frames", str(video_name))
for idx in frame_indices:
for fmt in (f"{int(idx):04d}.png", f"{int(idx)}.png"):
p = os.path.join(base, fmt)
if os.path.isfile(p):
out.append(Image.open(p).convert("RGB").resize((w, h)))
break
else:
return None
return out
def seed_for_sample(base_seed: int, video_name, start_frame) -> int:
"""与 (video_name, start_frame) 一一对应的确定性 seed,使同一样本在 loop_traj_fov_gen 与 evals_ep0 等不同调用下结果一致(避免因 idx/rank 不同导致偏移)。"""
try:
h = hash((str(video_name), int(start_frame)))
return base_seed + (h & 0x7FFFFFFF) % 100000
except Exception:
return base_seed
def sample_random_frame_from_dataset(dataset_base, w, h, seed, metadata_path=None):
"""从训练数据中随机采样一帧(用于首 chunk 的 I2V context,放右侧)。返回 PIL 或 None。"""
import csv
candidates = []
if metadata_path and os.path.isfile(metadata_path):
try:
with open(metadata_path, "r", encoding="utf-8") as f:
reader = csv.DictReader(f)
for row in reader:
vn = row.get("video_name", "").strip()
sf = row.get("start_frame", "")
if vn and str(sf).strip():
try:
candidates.append((vn, int(sf)))
except ValueError:
continue
except Exception:
pass
if not candidates:
frames_dir = os.path.join(dataset_base, "frames")
if os.path.isdir(frames_dir):
for vn in sorted(os.listdir(frames_dir)):
vd = os.path.join(frames_dir, vn)
if not os.path.isdir(vd):
continue
pngs = [f for f in os.listdir(vd) if f.endswith(".png")]
if not pngs:
continue
for p in pngs[:3]:
try:
frame_idx = int(os.path.splitext(p)[0])
candidates.append((vn, frame_idx))
except ValueError:
continue
if not candidates:
return None
rng = random.Random(seed)
vn, frame_idx = rng.choice(candidates)
img_p = os.path.join(dataset_base, "frames", vn, f"{frame_idx:04d}.png")
if not os.path.isfile(img_p):
img_p = os.path.join(dataset_base, "frames", vn, f"{frame_idx}.png")
if os.path.isfile(img_p):
return Image.open(img_p).convert("RGB").resize((w, h))
return None
def _angle_distance_deg(a: float, b: float) -> float:
"""Smallest absolute yaw distance in degrees."""
return abs((float(a) - float(b) + 180.0) % 360.0 - 180.0)
def fov_history_context_from_generated_frames(frames_list, yaw_list, K: int, target_yaw: float):
"""Select replay context by yaw/FOV proxy from generated history.
The first context frame remains the most recent frame for short-term continuity.
The remaining frames are selected from generated history by closest world-yaw
distance to the next chunk's midpoint yaw. Context actions are handled by the
caller and intentionally remain unchanged in this first version.
"""
n = min(len(frames_list), len(yaw_list))
if n <= 0 or K <= 0:
return [], []
if K == 1 or n == 1:
return [frames_list[n - 1]], [n - 1]
forced_idx = n - 1
need = min(int(K) - 1, n - 1)
scored = []
for idx in range(n - 1):
dist = _angle_distance_deg(yaw_list[idx], target_yaw)
# Prefer recent frames as a tie-breaker while keeping FOV/yaw match primary.
scored.append((dist, -idx, idx))
scored.sort()
selected = [idx for _dist, _neg_idx, idx in scored[:need]]
return [frames_list[forced_idx]] + [frames_list[i] for i in selected], [forced_idx] + selected
def _world_yaw_rt(yaw: float):
return compute_rotation_list([0.0, 0.0, 0.0, float(yaw)])
def context_actions_from_world_yaws(selected_yaws, ref_yaw: float):
"""Convert selected world-yaw RTs into RTs relative to the next chunk start."""
from src.model_training.fov_retrieval import convert_rt_to_relative
ref_rt = _world_yaw_rt(ref_yaw)
selected_rts = [_world_yaw_rt(y) for y in selected_yaws]
return convert_rt_to_relative(selected_rts, ref_rt)
def trim_continuation_first_frame(chunk_frames_list, yaw_history, chunk_frames=81):
"""与训练对齐:训练时 context[0]=target[0](同一帧),续段首帧=上一段末帧。拼接时去掉续段第 0 帧避免重复/跳帧。
返回 (trimmed_chunk_frames_list, trimmed_yaw):chunk0 保留 81 帧,chunk1 及之后只保留 [1:81] 共 80 帧。"""
if not chunk_frames_list or len(chunk_frames_list) <= 1:
return chunk_frames_list, yaw_history
trimmed_chunks = [list(chunk_frames_list[0])]
trimmed_yaw = []
offset = 0
for i, gen_frames in enumerate(chunk_frames_list):
n_use = len(gen_frames) if i == 0 else max(0, len(gen_frames) - 1)
if i == 0:
trimmed_yaw.extend(yaw_history[offset : offset + n_use])
else:
trimmed_yaw.extend(yaw_history[offset + 1 : offset + 1 + n_use])
offset += len(gen_frames)
if i > 0 and len(gen_frames) > 1:
trimmed_chunks.append(list(gen_frames[1:]))
elif i > 0:
trimmed_chunks.append([])
return trimmed_chunks, trimmed_yaw
def build_right_panel_and_yaw(chunk_frames_list, yaw_history, context_frames_per_chunk, chunk_frames=81, w=640, h=352, frame_to_pil_fn=None):
"""从每段生成帧 + 每段后的 context 帧拼出右栏;并扩展 yaw 与右栏帧数一致(context 帧沿用上一 yaw)。
支持变长 chunk(如 trim 后首段 81、续段 80),按 yaw_history 顺序逐帧对齐。"""
fp = frame_to_pil_fn or _frame_to_pil
right_frames = []
yaw_extended = []
yaw_idx = 0
for i, gen_frames in enumerate(chunk_frames_list):
for j, f in enumerate(gen_frames):
right_frames.append(f)
yaw_extended.append(yaw_history[yaw_idx] if yaw_idx < len(yaw_history) else (yaw_history[-1] if yaw_history else 0.0))
yaw_idx += 1
ctx_list = context_frames_per_chunk[i] if i < len(context_frames_per_chunk) else []
last_yaw = yaw_extended[-1] if yaw_extended else 0.0
for ctx in ctx_list:
right_frames.append(ctx if isinstance(ctx, Image.Image) else fp(ctx, w, h))
yaw_extended.append(last_yaw)
return right_frames, yaw_extended
def _frame_to_pil(f, tw, th):
if hasattr(f, "convert") and hasattr(f, "resize"):
return f.convert("RGB").resize((tw, th))
if isinstance(f, np.ndarray):
if f.dtype != np.uint8:
f = (f * 255).astype(np.uint8) if f.max() <= 1.0 else f.astype(np.uint8)
return Image.fromarray(f).convert("RGB").resize((tw, th))
if isinstance(f, torch.Tensor):
fn = f.cpu().numpy()
if len(fn.shape) == 3 and fn.shape[0] == 3:
fn = fn.transpose(1, 2, 0)
fn = (fn * 255).clip(0, 255).astype(np.uint8) if fn.max() <= 1.0 else fn.clip(0, 255).astype(np.uint8)
return Image.fromarray(fn).convert("RGB").resize((tw, th))
return f
def run_one_chunk(
pipe,
prompt,
use_negative_prompt,
action_path=None,
cam_pose_actions=None,
context_latents=None,
num_context_frames=1,
context_actions_t=None,
chunk_frames=81,
h=352,
w=640,
seed=0,
sigma_shift=15.0,
num_inference_steps=50,
cfg_scale=5.0,
inference_noise_level=0.0,
omit_context_actions=False, # kept for backward compat, no longer used
log_prefix="[Loop]",
**_extra,
):
"""Generate one chunk. VWM-aligned action injection via cam_pose_actions (latent-frame-aligned 12-D RT)."""
device = pipe.device
kwargs_common = dict(
prompt=prompt,
negative_prompt=use_negative_prompt,
height=h, width=w, num_frames=chunk_frames,
num_inference_steps=num_inference_steps,
seed=seed,
cfg_scale=cfg_scale,
sigma_shift=sigma_shift,
denoising_strength=1.0,
)
if action_path is not None:
kwargs_common["action_path"] = action_path
elif cam_pose_actions is not None:
kwargs_common["cam_pose_actions"] = cam_pose_actions
has_action_mlp = hasattr(pipe.dit.blocks[0], 'action_mlp') if len(pipe.dit.blocks) > 0 else False
if (action_path or cam_pose_actions is not None) and not has_action_mlp:
print(f"{log_prefix} 警告: dit.blocks[0] 无 action_mlp,action 可能未注入(请确认 ckpt 含 action_mlp 权重)", flush=True)
print(f"{log_prefix} run_one_chunk sigma_shift={sigma_shift} steps={num_inference_steps} (ctx={num_context_frames}) inference_noise={inference_noise_level} context_mem={context_latents is not None} action_mlp={has_action_mlp}")
if context_latents is not None:
pipe_kw = dict(
**kwargs_common,
enable_context_memory=True,
context_latents=context_latents,
num_context_frames=num_context_frames,
context_position="suffix",
cfg_target_only=True,
inference_noise_level=inference_noise_level,
)
if context_actions_t is not None:
pipe_kw["context_actions"] = context_actions_t
with torch.no_grad():
vid = pipe(**pipe_kw)
else:
with torch.no_grad():
vid = pipe(**kwargs_common, enable_context_memory=False)
return vid if isinstance(vid, list) else [vid]
def run_roundtrip_2chunk(
pipe,
dataset_base,
output_dir,
video_name,
start_frame,
deg=30.0,
chunk_frames=81,
context_frames=1,
h=352,
w=640,
sigma_shift=15.0,
num_inference_steps=50,
cfg_scale=5.0,
seed=42,
inference_noise_level=0.0,
metadata_path=None,
keep_action_jsons=False,
omit_context_actions=True,
):
"""2 chunk 折返:匀速 顺时针 deg° 再 逆时针 deg°(回起点)。"""
print(f"[Roundtrip] 匀速 顺时针{deg}° then 逆时针{deg}° sigma_shift={sigma_shift} omit_ctx_act={omit_context_actions}")
json_file = os.path.join(dataset_base, "jsons", f"{video_name}.json")
prompt = irc.load_prompt_for_video(dataset_base, video_name) or "A scene."
use_negative_prompt = getattr(irc, "DEFAULT_NEGATIVE_PROMPT", "oversaturated colors, overexposed, static, blurry details")
actions_ch1, actions_ch2 = build_action_cw_then_ccw(deg, chunk_frames)
subdir = os.path.join(output_dir, f"loop_{video_name}_start{start_frame}")
os.makedirs(subdir, exist_ok=True)
path_ch1 = os.path.join(subdir, "_ch1_cw.json")
path_ch2 = os.path.join(subdir, "_ch2_ccw.json")
with open(path_ch1, "w") as f:
json.dump(actions_ch1, f, indent=2)
with open(path_ch2, "w") as f:
json.dump(actions_ch2, f, indent=2)
print(f"[Roundtrip] Chunk1 CW 0->-{deg}° Chunk2 CCW 0->+{deg}° (匀速)")
denom = max(1, chunk_frames - 1)
yaw_history = [-(i / denom) * deg for i in range(chunk_frames)] # chunk1: 0 -> -deg
last_y1 = -deg
yaw_history += [last_y1 + (i / denom) * deg for i in range(chunk_frames)] # chunk2: -deg -> 0
chunk_frames_list = []
context_frames_per_chunk = []
# Chunk 1: context = 当前样本首帧(与 Replay 一致,同视频同场景),避免随机帧导致糊;若无则回退随机帧
ctx_pil_0 = load_sample_first_frame(dataset_base, video_name, start_frame, w, h)
used_sample_first = ctx_pil_0 is not None
if ctx_pil_0 is None:
ctx_pil_0 = sample_random_frame_from_dataset(dataset_base, w, h, seed, metadata_path)
if ctx_pil_0 is not None:
print(f"[Roundtrip] Chunk1 context: {'sample first frame ' + str((video_name, start_frame)) if used_sample_first else 'random frame (fallback)'}")
if ctx_pil_0 is not None:
ctx_pil_0 = [ctx_pil_0]
pipe.load_models_to_device(["vae"])
with torch.no_grad():
ctx_latents_0 = encode_context_frames_per_frame(pipe, ctx_pil_0, pipe.device)
identity_rt = [0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 1.0]
ctx_actions_t_0 = torch.tensor([identity_rt], dtype=torch.float32)
frames_ch1 = run_one_chunk(
pipe, prompt, use_negative_prompt, path_ch1,
context_latents=ctx_latents_0,
num_context_frames=ctx_latents_0.shape[2],
context_actions_t=ctx_actions_t_0,
chunk_frames=chunk_frames, h=h, w=w, seed=seed,
sigma_shift=sigma_shift, num_inference_steps=num_inference_steps,
cfg_scale=cfg_scale, inference_noise_level=inference_noise_level,
omit_context_actions=omit_context_actions,
)
context_frames_per_chunk.append([ctx_pil_0[0]])
else:
frames_ch1 = run_one_chunk(
pipe, prompt, use_negative_prompt, path_ch1,
chunk_frames=chunk_frames, h=h, w=w, seed=seed,
sigma_shift=sigma_shift, num_inference_steps=num_inference_steps, cfg_scale=cfg_scale,
omit_context_actions=omit_context_actions,
)
context_frames_per_chunk.append([_frame_to_pil(frames_ch1[-1], w, h)])
chunk_frames_list.append(frames_ch1)
# Chunk 2: condition = 上一 chunk,顺序与训练一致 [last_frame, ctx1, ...] 紧挨噪声后
n_ctx = min(context_frames, len(frames_ch1))
prev_frames = replay_context_from_generated_frames(frames_ch1, n_ctx)
prev_pil = [_frame_to_pil(f, w, h) for f in prev_frames]
identity_rt = [0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 1.0]
pipe.load_models_to_device(["vae"])
with torch.no_grad():
context_latents = encode_context_frames_per_frame(pipe, prev_pil, pipe.device)
num_ctx_tokens = context_latents.shape[2]
context_actions_t = torch.tensor([identity_rt] * num_ctx_tokens, dtype=torch.float32)
print(f"[Loop] continuation chunk: len(prev_pil)={len(prev_pil)} num_ctx_tokens={num_ctx_tokens} (应等于 ctx)")
frames_ch2 = run_one_chunk(
pipe, prompt, use_negative_prompt, path_ch2,
context_latents=context_latents,
num_context_frames=num_ctx_tokens,
context_actions_t=context_actions_t,
chunk_frames=chunk_frames, h=h, w=w, seed=seed + 1,
sigma_shift=sigma_shift, num_inference_steps=num_inference_steps,
cfg_scale=cfg_scale, inference_noise_level=inference_noise_level,
omit_context_actions=omit_context_actions,
)
chunk_frames_list.append(frames_ch2)
context_frames_per_chunk.append(list(prev_pil))
if not keep_action_jsons:
for p in [path_ch1, path_ch2]:
if os.path.exists(p):
try:
os.remove(p)
except Exception:
pass
return chunk_frames_list, yaw_history, context_frames_per_chunk
def run_single_chunk_rotation(
pipe,
dataset_base,
output_dir,
video_name,
start_frame,
deg=45.0,
clockwise=False,
chunk_frames=81,
h=352,
w=640,
sigma_shift=15.0,
num_inference_steps=50,
cfg_scale=5.0,
seed=42,
inference_noise_level=0.0,
metadata_path=None,
keep_action_jsons=False,
sampling_action_dir=None,
omit_context_actions=True,
context_image_path=None,
):
"""旋转原子操作:单 chunk 左转(CCW)或右转(CW) deg°。优先使用与训练采样相同的 action JSON。omit_context_actions 与训练 ctx 设置对齐。context_image_path:泛化检查时用指定图片作为首帧 context。"""
label = "右转(CW)" if clockwise else "左转(CCW)"
print(f"[SingleChunk] 1 chunk {label} {deg}° sigma_shift={sigma_shift}")
prompt = irc.load_prompt_for_video(dataset_base, video_name) or "A scene."
use_negative_prompt = getattr(irc, "DEFAULT_NEGATIVE_PROMPT", "oversaturated colors, overexposed, static, blurry details")
# 与采样对齐:45°、81 帧时优先用采样同款 JSON(generate_rotation_actions.py 生成),注入方式一致
path_ch = None
if deg == 45.0 and chunk_frames == 81:
action_dir = sampling_action_dir if sampling_action_dir else _script_dir
if clockwise:
candidate = os.path.join(action_dir, "action_rotation_right_45.json")
else:
candidate = os.path.join(action_dir, "action_rotation_left_45.json")
if os.path.isfile(candidate):
path_ch = candidate
print(f"[SingleChunk] 使用采样同款 action: {os.path.basename(path_ch)} (与训练采样注入一致)", flush=True)
if path_ch is None:
actions, yaw_chunk = build_action_chunk(deg, clockwise, chunk_frames)
yaw0 = _yaw_deg_from_rt(actions["0"])
yaw_last = _yaw_deg_from_rt(actions[str(chunk_frames - 1)])
print(f"[SingleChunk] action 首帧 yaw={yaw0:.1f}° 末帧 yaw={yaw_last:.1f}° (预期: 左转 0→+{deg}° 右转 0→-{deg}°)", flush=True)
subdir = os.path.join(output_dir, f"loop_{video_name}_start{start_frame}")
os.makedirs(subdir, exist_ok=True)
suffix = "cw_right" if clockwise else "ccw_left"
path_ch = os.path.join(subdir, f"_single_{suffix}.json")
with open(path_ch, "w") as f:
json.dump(actions, f, indent=2)
yaw_history = list(yaw_chunk)
delete_path_after = not keep_action_jsons
else:
with open(path_ch, "r") as f:
actions = json.load(f)
yaw_history = [_yaw_deg_from_rt(actions.get(str(i), [0] * 12)) for i in range(min(chunk_frames, len(actions)))]
if len(yaw_history) < chunk_frames:
yaw_history += [yaw_history[-1]] * (chunk_frames - len(yaw_history))
delete_path_after = False
subdir = os.path.join(output_dir, f"loop_{video_name}_start{start_frame}")
os.makedirs(subdir, exist_ok=True)
yaw_history = yaw_history[:chunk_frames] if len(yaw_history) > chunk_frames else yaw_history
chunk_frames_list = []
context_frames_per_chunk = []
if context_image_path and os.path.isfile(context_image_path):
ctx_pil_0 = Image.open(context_image_path).convert("RGB").resize((w, h))
print(f"[SingleChunk] 使用指定 context 图: {context_image_path}", flush=True)
else:
ctx_pil_0 = load_sample_first_frame(dataset_base, video_name, start_frame, w, h)
if ctx_pil_0 is None:
ctx_pil_0 = sample_random_frame_from_dataset(dataset_base, w, h, seed, metadata_path)
if ctx_pil_0 is not None:
ctx_pil_0 = [ctx_pil_0]
pipe.load_models_to_device(["vae"])
with torch.no_grad():
ctx_latents_0 = encode_context_frames_per_frame(pipe, ctx_pil_0, pipe.device)
identity_rt = [0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 1.0]
ctx_actions_t_0 = torch.tensor([identity_rt], dtype=torch.float32)
frames_ch = run_one_chunk(
pipe, prompt, use_negative_prompt, path_ch,
context_latents=ctx_latents_0,
num_context_frames=ctx_latents_0.shape[2],
context_actions_t=ctx_actions_t_0,
chunk_frames=chunk_frames, h=h, w=w, seed=seed,
sigma_shift=sigma_shift, num_inference_steps=num_inference_steps,
cfg_scale=cfg_scale, inference_noise_level=inference_noise_level,
omit_context_actions=omit_context_actions,
)
context_frames_per_chunk.append([ctx_pil_0[0]])
else:
frames_ch = run_one_chunk(
pipe, prompt, use_negative_prompt, path_ch,
chunk_frames=chunk_frames, h=h, w=w, seed=seed,
sigma_shift=sigma_shift, num_inference_steps=num_inference_steps, cfg_scale=cfg_scale,
omit_context_actions=omit_context_actions,
)
context_frames_per_chunk.append([_frame_to_pil(frames_ch[-1], w, h)])
chunk_frames_list.append(frames_ch)
if delete_path_after and os.path.exists(path_ch):
try:
os.remove(path_ch)
except Exception:
pass
return chunk_frames_list, yaw_history, context_frames_per_chunk
def run_left_right_2chunk(
pipe,
dataset_base,
output_dir,
video_name,
start_frame,
deg=45.0,
chunk_frames=81,
context_frames=1,
h=352,
w=640,
sigma_shift=15.0,
num_inference_steps=50,
cfg_scale=5.0,
seed=42,
inference_noise_level=0.0,
metadata_path=None,
keep_action_jsons=False,
sampling_action_dir=None,
omit_context_actions=True,
):
"""回环:1 chunk 左转(CCW) deg°,1 chunk 右转(CW) deg°。45°×81 时优先用采样同款 left_45/right_45.json。"""
print(f"[LeftRight 2chunk] 左转(CCW){deg}° then 右转(CW){deg}° sigma_shift={sigma_shift} omit_ctx_act={omit_context_actions} context_frames={context_frames} (续段 ctx 数)")
prompt = irc.load_prompt_for_video(dataset_base, video_name) or "A scene."
use_negative_prompt = getattr(irc, "DEFAULT_NEGATIVE_PROMPT", "oversaturated colors, overexposed, static, blurry details")
subdir = os.path.join(output_dir, f"loop_{video_name}_start{start_frame}")
os.makedirs(subdir, exist_ok=True)
action_dir = sampling_action_dir if sampling_action_dir else _script_dir
path_ch1 = path_ch2 = None
delete_ch1 = delete_ch2 = True
if deg == 45.0 and chunk_frames == 81:
p_left = os.path.join(action_dir, "action_rotation_left_45.json")
p_right = os.path.join(action_dir, "action_rotation_right_45.json")
if os.path.isfile(p_left) and os.path.isfile(p_right):
path_ch1, path_ch2 = p_left, p_right
delete_ch1 = delete_ch2 = False
print(f"[LeftRight 2chunk] 使用采样同款 action: left_45 + right_45 (与训练采样注入一致)", flush=True)
if path_ch1 is None:
actions_ccw, actions_cw = build_action_ccw_cw(deg, chunk_frames)
path_ch1 = os.path.join(subdir, "_ch1_ccw_left.json")
path_ch2 = os.path.join(subdir, "_ch2_cw_right.json")
with open(path_ch1, "w") as f:
json.dump(actions_ccw, f, indent=2)
with open(path_ch2, "w") as f:
json.dump(actions_cw, f, indent=2)
delete_ch1 = delete_ch2 = not keep_action_jsons
if not delete_ch1 and path_ch1 and path_ch2:
with open(path_ch1, "r") as f:
a1 = json.load(f)
with open(path_ch2, "r") as f:
a2 = json.load(f)
yaw_left = [_yaw_deg_from_rt(a1.get(str(i), [0] * 12)) for i in range(min(chunk_frames, len(a1)))]
yaw_right = [_yaw_deg_from_rt(a2.get(str(i), [0] * 12)) for i in range(min(chunk_frames, len(a2)))]
if len(yaw_left) < chunk_frames:
yaw_left += [yaw_left[-1]] * (chunk_frames - len(yaw_left))
if len(yaw_right) < chunk_frames:
yaw_right += [yaw_right[-1]] * (chunk_frames - len(yaw_right))
base = yaw_left[-1] if yaw_left else 45.0
yaw_history = yaw_left[:chunk_frames] + [base + yaw_right[i] for i in range(chunk_frames)]
else:
denom = max(1, chunk_frames - 1)
yaw_history = [(i / denom) * deg for i in range(chunk_frames)]
yaw_history += [deg - (i / denom) * deg for i in range(chunk_frames)]
chunk_frames_list = []
context_frames_per_chunk = []
ctx_pil_0 = load_sample_first_frame(dataset_base, video_name, start_frame, w, h)
used_sample_first = ctx_pil_0 is not None
if ctx_pil_0 is None:
ctx_pil_0 = sample_random_frame_from_dataset(dataset_base, w, h, seed, metadata_path)
if ctx_pil_0 is not None:
ctx_pil_0 = [ctx_pil_0]
pipe.load_models_to_device(["vae"])
with torch.no_grad():
ctx_latents_0 = encode_context_frames_per_frame(pipe, ctx_pil_0, pipe.device)
identity_rt = [0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 1.0]
ctx_actions_t_0 = torch.tensor([identity_rt], dtype=torch.float32)
frames_ch1 = run_one_chunk(
pipe, prompt, use_negative_prompt, path_ch1,
context_latents=ctx_latents_0,
num_context_frames=ctx_latents_0.shape[2],
context_actions_t=ctx_actions_t_0,
chunk_frames=chunk_frames, h=h, w=w, seed=seed,
sigma_shift=sigma_shift, num_inference_steps=num_inference_steps,
cfg_scale=cfg_scale, inference_noise_level=inference_noise_level,
omit_context_actions=omit_context_actions,
)
context_frames_per_chunk.append([ctx_pil_0[0]])
else:
frames_ch1 = run_one_chunk(
pipe, prompt, use_negative_prompt, path_ch1,
chunk_frames=chunk_frames, h=h, w=w, seed=seed,
sigma_shift=sigma_shift, num_inference_steps=num_inference_steps, cfg_scale=cfg_scale,
omit_context_actions=omit_context_actions,
)
context_frames_per_chunk.append([_frame_to_pil(frames_ch1[-1], w, h)])
chunk_frames_list.append(frames_ch1)
n_ctx = min(context_frames, len(frames_ch1))
prev_frames = replay_context_from_generated_frames(frames_ch1, n_ctx)
prev_pil = [_frame_to_pil(f, w, h) for f in prev_frames]
identity_rt = [0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 1.0]
pipe.load_models_to_device(["vae"])
with torch.no_grad():
context_latents = encode_context_frames_per_frame(pipe, prev_pil, pipe.device)
num_ctx_tokens = context_latents.shape[2]
context_actions_t = torch.tensor([identity_rt] * num_ctx_tokens, dtype=torch.float32)
print(f"[Loop] continuation chunk: len(prev_pil)={len(prev_pil)} num_ctx_tokens={num_ctx_tokens} (应等于 ctx)")
frames_ch2 = run_one_chunk(
pipe, prompt, use_negative_prompt, path_ch2,
context_latents=context_latents,
num_context_frames=num_ctx_tokens,
context_actions_t=context_actions_t,
chunk_frames=chunk_frames, h=h, w=w, seed=seed + 1,
sigma_shift=sigma_shift, num_inference_steps=num_inference_steps,
cfg_scale=cfg_scale, inference_noise_level=inference_noise_level,
omit_context_actions=omit_context_actions,
)
chunk_frames_list.append(frames_ch2)
context_frames_per_chunk.append(list(prev_pil))
if delete_ch1 and path_ch1 and os.path.exists(path_ch1):
try:
os.remove(path_ch1)
except Exception:
pass
if delete_ch2 and path_ch2 and os.path.exists(path_ch2):
try:
os.remove(path_ch2)
except Exception:
pass
return chunk_frames_list, yaw_history, context_frames_per_chunk
def run_translation_4chunk(
pipe,
dataset_base,
output_dir,
video_name,
start_frame,
direction="forward",
translation_delta=0.1,
chunk_frames=81,
context_frames=1,
h=352,
w=640,
sigma_shift=15.0,
num_inference_steps=50,
cfg_scale=5.0,
seed=42,
inference_noise_level=0.0,
metadata_path=None,
omit_context_actions=True,
):
"""仅纯平移 4chunk:同一方向(forward/backward/left/right)连续 4 段,R=I,无旋转。仅用于 --run_atomic_translation_4chunk。"""
_prefix = "[atomic_translation_4chunk]"
print(f"{_prefix} 纯平移 direction={direction} delta={translation_delta} sigma_shift={sigma_shift} omit_ctx_act={omit_context_actions} context_frames={context_frames} (续段 ctx 数)")
prompt = irc.load_prompt_for_video(dataset_base, video_name) or "A scene."
use_negative_prompt = getattr(irc, "DEFAULT_NEGATIVE_PROMPT", "oversaturated colors, overexposed, static, blurry details")
yaw_history = []
chunk_frames_list = []
context_frames_per_chunk = []
subdir = os.path.join(output_dir, f"loop_{video_name}_start{start_frame}")
os.makedirs(subdir, exist_ok=True)
identity_rt = [0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 1.0]
context_latents = None
context_actions_t = None
for ch in range(4):
actions, yaw_chunk = build_action_translation_only(direction, translation_delta, chunk_frames)
path_ch = os.path.join(subdir, f"_ch{ch}_trans_{direction}.json")
with open(path_ch, "w") as f:
json.dump(actions, f, indent=2)
try:
if ch == 0:
ctx_pil_0 = load_sample_first_frame(dataset_base, video_name, start_frame, w, h)
if ctx_pil_0 is None:
ctx_pil_0 = sample_random_frame_from_dataset(dataset_base, w, h, seed, metadata_path)
if ctx_pil_0 is not None:
ctx_pil_0 = [ctx_pil_0]
pipe.load_models_to_device(["vae"])
with torch.no_grad():
context_latents = encode_context_frames_per_frame(pipe, ctx_pil_0, pipe.device)
context_actions_t = torch.tensor([identity_rt], dtype=torch.float32)
context_frames_per_chunk.append([ctx_pil_0[0]])
else:
context_frames_per_chunk.append([])
if context_latents is not None and context_actions_t is not None:
frames_ch = run_one_chunk(
pipe, prompt, use_negative_prompt, path_ch,
context_latents=context_latents,
num_context_frames=context_latents.shape[2],
context_actions_t=context_actions_t,
chunk_frames=chunk_frames, h=h, w=w, seed=seed + ch,
sigma_shift=sigma_shift, num_inference_steps=num_inference_steps,
cfg_scale=cfg_scale, inference_noise_level=inference_noise_level,
omit_context_actions=omit_context_actions,
log_prefix=_prefix,
)
else:
frames_ch = run_one_chunk(
pipe, prompt, use_negative_prompt, path_ch,
chunk_frames=chunk_frames, h=h, w=w, seed=seed + ch,
sigma_shift=sigma_shift, num_inference_steps=num_inference_steps, cfg_scale=cfg_scale,
omit_context_actions=omit_context_actions,
log_prefix=_prefix,
)
if ch == 0 and not context_frames_per_chunk[0]:
context_frames_per_chunk[0] = [_frame_to_pil(frames_ch[-1], w, h)]
for y in yaw_chunk:
yaw_history.append(y)
chunk_frames_list.append(frames_ch)
if ch < 3:
n_ctx = min(context_frames, len(frames_ch))
prev_frames = context_frames_for_next_chunk(frames_ch, n_ctx) if n_ctx else [frames_ch[-1]]
prev_pil = [_frame_to_pil(f, w, h) for f in prev_frames]
context_frames_per_chunk.append(list(prev_pil))
pipe.load_models_to_device(["vae"])
with torch.no_grad():
context_latents = encode_context_frames_per_frame(pipe, prev_pil, pipe.device)
num_ctx_tokens = context_latents.shape[2]
context_actions_t = torch.tensor([identity_rt] * num_ctx_tokens, dtype=torch.float32)
print(f"{_prefix} continuation chunk ch={ch}: len(prev_pil)={len(prev_pil)} num_ctx_tokens={num_ctx_tokens} (应等于 ctx)")
finally:
if os.path.exists(path_ch):
try:
os.remove(path_ch)
except Exception:
pass
return chunk_frames_list, yaw_history, context_frames_per_chunk
def run_left2_right2_4chunk(
pipe,
dataset_base,
output_dir,
video_name,
start_frame,
deg_per_chunk=45.0,
chunk_frames=81,
context_frames=1,
h=352,
w=640,
sigma_shift=15.0,
num_inference_steps=50,
cfg_scale=5.0,
seed=42,
inference_noise_level=0.0,
metadata_path=None,
sampling_action_dir=None,
omit_context_actions=True,
multi_ctx_all_history=False,
fov_history_context=False,
fov_last_target=False,
fov_context_rt=False,
):
"""回环:先左转 2 chunk(各 deg° CCW),再右转 2 chunk(各 deg° CW)。
默认行为:与 2chunk 回环一致,续段 context 仅来自上一 chunk 的末尾若干帧。
当 multi_ctx_all_history=True 时:续段 context 第一帧必须为上一 chunk 的最后一帧;
其余 (ctx-1) 帧从「前面所有已生成帧」均匀采样(保持时间序),用于记忆机制对比。
当 fov_history_context=True 时:续段 context 第一帧仍为上一 chunk 的最后一帧;
其余 (ctx-1) 帧按下一 chunk 中点 world-yaw,从所有已生成历史帧中检索最接近的帧。
当 fov_last_target=True 时:仅第 4 个 chunk 的检索 target 改用该 chunk 末尾 world-yaw。
当 fov_context_rt=True 时:FOV/yaw 检索出的 context actions 转成相对下一 chunk 起始 world-yaw 的 RT。
"""
print(f"[Left2Right2 4chunk] 左转2chunk then 右转2chunk sigma_shift={sigma_shift} omit_ctx_act={omit_context_actions} context_frames={context_frames} (续段 ctx 数)")
prompt = irc.load_prompt_for_video(dataset_base, video_name) or "A scene."
use_negative_prompt = getattr(irc, "DEFAULT_NEGATIVE_PROMPT", "oversaturated colors, overexposed, static, blurry details")
yaw_history = []
chunk_frames_list = []
context_frames_per_chunk = []
all_prev_frames = []
subdir = os.path.join(output_dir, f"loop_{video_name}_start{start_frame}")
os.makedirs(subdir, exist_ok=True)
identity_rt = [0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 1.0]
cumulative_yaw = 0.0
context_latents = None
context_actions_t = None
# 45°×81 时直接使用采样同款 JSON,多 chunk 多次导入同一文件即可
action_dir = sampling_action_dir if sampling_action_dir else _script_dir
path_left = os.path.join(action_dir, "action_rotation_left_45.json")
path_right = os.path.join(action_dir, "action_rotation_right_45.json")
use_sampling_files = (deg_per_chunk == 45.0 and chunk_frames == 81 and
os.path.isfile(path_left) and os.path.isfile(path_right))
if use_sampling_files:
print(f"[Left2Right2 4chunk] 使用采样同款 action: left_45 x2 + right_45 x2 (与训练采样注入一致)", flush=True)
with open(path_left, "r") as f:
_actions_left = json.load(f)
with open(path_right, "r") as f:
_actions_right = json.load(f)
_yaw_left = [_yaw_deg_from_rt(_actions_left.get(str(i), [0] * 12)) for i in range(min(chunk_frames, len(_actions_left)))]
_yaw_right = [_yaw_deg_from_rt(_actions_right.get(str(i), [0] * 12)) for i in range(min(chunk_frames, len(_actions_right)))]
if len(_yaw_left) < chunk_frames:
_yaw_left += [_yaw_left[-1]] * (chunk_frames - len(_yaw_left))
if len(_yaw_right) < chunk_frames:
_yaw_right += [_yaw_right[-1]] * (chunk_frames - len(_yaw_right))
for ch in range(4):
clockwise = ch >= 2
if use_sampling_files:
path_ch = path_right if clockwise else path_left
else:
actions, yaw_chunk = build_action_chunk(deg_per_chunk, clockwise, chunk_frames)
path_ch = os.path.join(subdir, f"_ch{ch}_left2right2.json")
with open(path_ch, "w") as f:
json.dump(actions, f, indent=2)
# 与 2chunk 一致:ch=0 用首帧 context;ch>=1 用上一 chunk 输出做 context(上一轮末尾已写好 context_latents/context_actions_t)
if ch == 0:
ctx_pil_0 = load_sample_first_frame(dataset_base, video_name, start_frame, w, h)
if ctx_pil_0 is None:
ctx_pil_0 = sample_random_frame_from_dataset(dataset_base, w, h, seed, metadata_path)
if ctx_pil_0 is not None:
ctx_pil_0 = [ctx_pil_0]
pipe.load_models_to_device(["vae"])
with torch.no_grad():
context_latents = encode_context_frames_per_frame(pipe, ctx_pil_0, pipe.device)
context_actions_t = torch.tensor([identity_rt], dtype=torch.float32)
context_frames_per_chunk.append([ctx_pil_0[0]])
else:
context_frames_per_chunk.append([])
if context_latents is not None and context_actions_t is not None:
frames_ch = run_one_chunk(
pipe, prompt, use_negative_prompt, path_ch,
context_latents=context_latents,
num_context_frames=context_latents.shape[2],
context_actions_t=context_actions_t,
chunk_frames=chunk_frames, h=h, w=w, seed=seed + ch,
sigma_shift=sigma_shift, num_inference_steps=num_inference_steps,
cfg_scale=cfg_scale, inference_noise_level=inference_noise_level,
omit_context_actions=omit_context_actions,
)
else:
frames_ch = run_one_chunk(
pipe, prompt, use_negative_prompt, path_ch,
chunk_frames=chunk_frames, h=h, w=w, seed=seed + ch,
sigma_shift=sigma_shift, num_inference_steps=num_inference_steps, cfg_scale=cfg_scale,
omit_context_actions=omit_context_actions,
)
if ch == 0 and not context_frames_per_chunk[0]:
context_frames_per_chunk[0] = [_frame_to_pil(frames_ch[-1], w, h)]
if use_sampling_files:
yaw_chunk = _yaw_right if clockwise else _yaw_left
else:
yaw_chunk = list(yaw_chunk)
for y in yaw_chunk:
yaw_history.append(cumulative_yaw + y)
cumulative_yaw = yaw_history[-1]
chunk_frames_list.append(frames_ch)
all_prev_frames.extend(frames_ch)
# 续段 context:
# - 默认:与 2chunk 一致,仅用上一 chunk 的 last n 帧
# - fov_history_context=True:第一帧必须为上一 chunk 最后一帧;其余 (ctx-1) 帧按下一 chunk 中点 world-yaw 做 FOV/yaw proxy 检索。
# - multi_ctx_all_history=True:第一帧必须为上一 chunk 最后一帧;其余 (ctx-1) 帧从「前面所有帧」均匀采样,保持时间序。
if ch < 3:
next_context_actions = None
if context_frames <= 0:
prev_frames = [frames_ch[-1]]
elif fov_history_context and len(all_prev_frames) > 0:
next_ch = ch + 1
next_clockwise = next_ch >= 2
if use_sampling_files:
next_yaw_chunk = _yaw_right if next_clockwise else _yaw_left
else:
_unused_actions, next_yaw_chunk = build_action_chunk(deg_per_chunk, next_clockwise, chunk_frames)
target_idx = min(max(0, chunk_frames // 2), len(next_yaw_chunk) - 1)
if fov_last_target and next_ch == 3:
target_idx = len(next_yaw_chunk) - 1
target_mid_yaw = cumulative_yaw + float(next_yaw_chunk[target_idx])
prev_frames, picked_indices = fov_history_context_from_generated_frames(
all_prev_frames,
yaw_history,
context_frames,
target_mid_yaw,
)
picked_yaws = [float(yaw_history[i]) for i in picked_indices if i < len(yaw_history)]
if fov_context_rt and picked_yaws:
next_context_actions = context_actions_from_world_yaws(picked_yaws, cumulative_yaw)
print(
f"[Loop] fov_history_context ch={ch}->next={next_ch}: "
f"target_idx={target_idx} target_yaw={target_mid_yaw:.2f} picked_indices={picked_indices} "
f"picked_yaws={[round(y, 2) for y in picked_yaws]} "
f"context_rt={bool(next_context_actions)} ref_yaw={cumulative_yaw:.2f}",
flush=True,
)
elif multi_ctx_all_history and len(all_prev_frames) > 0:
total = len(all_prev_frames)
if total == 1 or context_frames == 1:
prev_frames = [all_prev_frames[-1]]
else:
# 第一帧 = 上一 chunk 最后一帧 (all_prev_frames[-1]);其余 (ctx-1) 帧从 all_prev_frames[0..total-2] 均匀采样
n_pick = min(context_frames - 1, total - 1)
if n_pick <= 0:
prev_frames = [all_prev_frames[-1]]
else:
pool_size = total - 1 # 可选下标 0..total-2(前面所有帧,不含已单独占位的最后一帧)
indices = []
for i in range(n_pick):
pos = int(round(i * (pool_size - 1) / max(n_pick - 1, 1)))
indices.append(max(0, min(pool_size - 1, pos)))
uniq_indices = sorted(set(indices))
while len(uniq_indices) < n_pick:
uniq_indices.append(pool_size - 1 if pool_size > 0 else 0)
uniq_indices = sorted(uniq_indices)[:n_pick]
chosen = [all_prev_frames[i] for i in uniq_indices]
prev_frames = [all_prev_frames[-1]] + chosen
else:
n_ctx = min(context_frames, len(frames_ch))
prev_frames = context_frames_for_next_chunk(frames_ch, n_ctx) if n_ctx else [frames_ch[-1]]
prev_pil = [_frame_to_pil(f, w, h) for f in prev_frames]
context_frames_per_chunk.append(list(prev_pil))
pipe.load_models_to_device(["vae"])
with torch.no_grad():
context_latents = encode_context_frames_per_frame(pipe, prev_pil, pipe.device)
num_ctx_tokens = context_latents.shape[2]
if next_context_actions is not None:
if len(next_context_actions) < num_ctx_tokens:
next_context_actions = next_context_actions + [next_context_actions[-1]] * (num_ctx_tokens - len(next_context_actions))
context_actions_t = torch.tensor(next_context_actions[:num_ctx_tokens], dtype=torch.float32)
else:
context_actions_t = torch.tensor([identity_rt] * num_ctx_tokens, dtype=torch.float32)
print(
f"[Loop] continuation chunk ch={ch}: len(prev_pil)={len(prev_pil)} num_ctx_tokens={num_ctx_tokens} "
f"(multi_ctx_all_history={multi_ctx_all_history} fov_history_context={fov_history_context} fov_context_rt={fov_context_rt})"
)
if not use_sampling_files and os.path.exists(path_ch):
try:
os.remove(path_ch)
except Exception:
pass
return chunk_frames_list, yaw_history, context_frames_per_chunk
def run_replay_4chunk(
pipe,
dataset_base,
output_dir,
video_name,
start_frame,
chunk_frames=81,
context_frames=1,
h=352,
w=640,
sigma_shift=15.0,
num_inference_steps=50,
cfg_scale=5.0,
seed=42,
inference_noise_level=0.0,
omit_context_actions=True,
):
"""4 chunk 参考训练:每段用同视频连续 4 段的 GT 轨迹,逐 chunk 对比。需 start_frame 起至少 4*81 帧。"""
print(f"[Replay 4chunk] {video_name} start{start_frame} (4 段 GT 轨迹) context_frames={context_frames} (续段 ctx 数)")
json_file = os.path.join(dataset_base, "jsons", f"{video_name}.json")
if not os.path.isfile(json_file):
return None
num_chunks = 4
need_frames = num_chunks * chunk_frames # 324
# 检查是否有足够帧(用 pose 存在性粗略判断)
try:
last_pose = irc.load_pose_rt(json_file, start_frame + need_frames - 1)
if last_pose is None or len(last_pose) < 12:
return None
except Exception:
return None
prompt = irc.load_prompt_for_video(dataset_base, video_name) or "A scene."
use_negative_prompt = getattr(irc, "DEFAULT_NEGATIVE_PROMPT", "oversaturated colors, overexposed, static, blurry details")
identity_rt = [0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 1.0]
subdir = os.path.join(output_dir, f"loop_{video_name}_start{start_frame}")
os.makedirs(subdir, exist_ok=True)
chunk_frames_list = []
context_frames_per_chunk = []
yaw_history = []
cumulative_yaw = 0.0
context_latents = None
context_actions_t = None
for ch in range(num_chunks):
seg_start = start_frame + ch * chunk_frames
ch1_actions = build_gt_trajectory_actions(dataset_base, video_name, seg_start, chunk_frames, json_file)
if ch1_actions is None:
return None
path_ch = os.path.join(subdir, f"_ch{ch}_replay.json")
with open(path_ch, "w") as f:
json.dump(ch1_actions, f, indent=2)
if ch == 0:
ctx_pil_0 = load_sample_first_frame(dataset_base, video_name, seg_start, w, h)
if ctx_pil_0 is None:
return None
ctx_pil_0 = [ctx_pil_0]
pipe.load_models_to_device(["vae"])
with torch.no_grad():
context_latents = encode_context_frames_per_frame(pipe, ctx_pil_0, pipe.device)
context_actions_t = torch.tensor([identity_rt], dtype=torch.float32)
context_frames_per_chunk.append([ctx_pil_0[0]])
else:
# 下一 chunk 的 condition = 上一 chunk 的 last n 帧(生成结果)
n_ctx = min(context_frames, len(chunk_frames_list[-1]))
prev_frames = context_frames_for_next_chunk(chunk_frames_list[-1], n_ctx) if n_ctx else [chunk_frames_list[-1][-1]]
prev_pil = [_frame_to_pil(f, w, h) for f in prev_frames]
if ch < num_chunks:
context_frames_per_chunk.append(list(prev_pil))
pipe.load_models_to_device(["vae"])
with torch.no_grad():
context_latents = encode_context_frames_per_frame(pipe, prev_pil, pipe.device)
context_actions_t = torch.tensor([identity_rt] * context_latents.shape[2], dtype=torch.float32)
print(f"[Loop] continuation chunk ch={ch}: len(prev_pil)={len(prev_pil)} num_ctx_tokens={context_latents.shape[2]} (应等于 ctx)")
frames_ch = run_one_chunk(
pipe, prompt, use_negative_prompt, path_ch,
context_latents=context_latents,
num_context_frames=context_latents.shape[2],
context_actions_t=context_actions_t,
chunk_frames=chunk_frames, h=h, w=w, seed=seed + ch * 100,
sigma_shift=sigma_shift, num_inference_steps=num_inference_steps,
cfg_scale=cfg_scale, inference_noise_level=inference_noise_level,
omit_context_actions=omit_context_actions,
)
for i in range(chunk_frames):
yaw_history.append(cumulative_yaw + _yaw_deg_from_rt(ch1_actions[str(i)]))
cumulative_yaw = yaw_history[-1]
chunk_frames_list.append(frames_ch)
if os.path.exists(path_ch):
try:
os.remove(path_ch)
except Exception:
pass
return chunk_frames_list, yaw_history, context_frames_per_chunk
def run_one_direction_4chunk(
pipe,
dataset_base,
output_dir,
video_name,
start_frame,
deg_per_chunk=15.0,
clockwise=True,
chunk_frames=81,
context_frames=1,
h=352,
w=640,
sigma_shift=15.0,
num_inference_steps=50,
cfg_scale=5.0,
seed=42,
inference_noise_level=0.0,
metadata_path=None,
omit_context_actions=True,
):
"""4 chunk 单向匀速:共 4*deg_per_chunk 度。omit_context_actions 与训练对齐。"""
print(f"[Onedir 4chunk] 匀速 总{4*deg_per_chunk:.0f}° {'CW' if clockwise else 'CCW'} sigma_shift={sigma_shift} omit_ctx_act={omit_context_actions} context_frames={context_frames} (续段 ctx 数)")
prompt = irc.load_prompt_for_video(dataset_base, video_name) or "A scene."
use_negative_prompt = getattr(irc, "DEFAULT_NEGATIVE_PROMPT", "oversaturated colors, overexposed, static, blurry details") # 与训练一致
yaw_history = []
chunk_frames_list = []
context_frames_per_chunk = []
subdir = os.path.join(output_dir, f"loop_{video_name}_start{start_frame}")
os.makedirs(subdir, exist_ok=True)
identity_rt = [0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 1.0]
cumulative_yaw_start = 0.0
context_latents = None
context_actions_t = None
for ch in range(4):
actions, yaw_chunk = build_action_chunk(deg_per_chunk, clockwise, chunk_frames)
path_ch = os.path.join(subdir, f"_ch{ch}_onedir.json")
with open(path_ch, "w") as f:
json.dump(actions, f, indent=2)
if ch == 0:
ctx_pil_0 = load_sample_first_frame(dataset_base, video_name, start_frame, w, h)
if ctx_pil_0 is None:
ctx_pil_0 = sample_random_frame_from_dataset(dataset_base, w, h, seed, metadata_path)
if ctx_pil_0 is not None:
ctx_pil_0 = [ctx_pil_0]
pipe.load_models_to_device(["vae"])
with torch.no_grad():
context_latents = encode_context_frames_per_frame(pipe, ctx_pil_0, pipe.device)
context_actions_t = torch.tensor([identity_rt], dtype=torch.float32)
context_frames_per_chunk.append([ctx_pil_0[0]])
else:
context_frames_per_chunk.append([])
if context_latents is not None and context_actions_t is not None:
frames_ch = run_one_chunk(
pipe, prompt, use_negative_prompt, path_ch,
context_latents=context_latents,
num_context_frames=context_latents.shape[2],
context_actions_t=context_actions_t,
chunk_frames=chunk_frames, h=h, w=w, seed=seed + ch,
sigma_shift=sigma_shift, num_inference_steps=num_inference_steps,
cfg_scale=cfg_scale, inference_noise_level=inference_noise_level,
omit_context_actions=omit_context_actions,
)
else:
frames_ch = run_one_chunk(
pipe, prompt, use_negative_prompt, path_ch,
chunk_frames=chunk_frames, h=h, w=w, seed=seed + ch,
sigma_shift=sigma_shift, num_inference_steps=num_inference_steps, cfg_scale=cfg_scale,
omit_context_actions=omit_context_actions,
)
if ch == 0 and not context_frames_per_chunk[0]:
context_frames_per_chunk[0] = [_frame_to_pil(frames_ch[-1], w, h)]
for y in yaw_chunk:
yaw_history.append(cumulative_yaw_start + y)
cumulative_yaw_start = yaw_history[-1]
chunk_frames_list.append(frames_ch)
# 与 2chunk 一致:仅当有下一 chunk 时用本 chunk 输出做 context(last n 帧 → 逐帧 VAE)
if ch < 3:
n_ctx = min(context_frames, len(frames_ch))
prev_frames = context_frames_for_next_chunk(frames_ch, n_ctx) if n_ctx else [frames_ch[-1]]
prev_pil = [_frame_to_pil(f, w, h) for f in prev_frames]
context_frames_per_chunk.append(list(prev_pil))
pipe.load_models_to_device(["vae"])
with torch.no_grad():
context_latents = encode_context_frames_per_frame(pipe, prev_pil, pipe.device)
num_ctx_tokens = context_latents.shape[2]
context_actions_t = torch.tensor([identity_rt] * num_ctx_tokens, dtype=torch.float32)
print(f"[Loop] next continuation chunk: len(prev_pil)={len(prev_pil)} num_ctx_tokens={num_ctx_tokens} (应等于 ctx)")
if os.path.exists(path_ch):
try:
os.remove(path_ch)
except Exception:
pass
return chunk_frames_list, yaw_history, context_frames_per_chunk
def run_random_4chunk(
pipe,
dataset_base,
output_dir,
video_name,
start_frame,
deg_min=15.0,
deg_max=30.0,
chunk_frames=81,
context_frames=1,
h=352,
w=640,
sigma_shift=15.0,
num_inference_steps=50,
cfg_scale=5.0,
seed=42,
inference_noise_level=0.0,
metadata_path=None,
omit_context_actions=True,
):
"""4 chunk 随机。首 chunk context=随机 1 帧;之后 context=上一 chunk last n。"""
prompt = irc.load_prompt_for_video(dataset_base, video_name) or "A scene."
use_negative_prompt = getattr(irc, "DEFAULT_NEGATIVE_PROMPT", "oversaturated colors, overexposed, static, blurry details") # 与训练一致
rng = random.Random(seed)
yaw_history = []
chunk_frames_list = []
context_frames_per_chunk = []
subdir = os.path.join(output_dir, f"loop_{video_name}_start{start_frame}")
os.makedirs(subdir, exist_ok=True)
identity_rt = [0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 1.0]
cumulative_yaw_start = 0.0
context_latents = None
context_actions_t = None
for ch in range(4):
deg = rng.uniform(deg_min, deg_max)
clockwise = rng.choice([True, False])
actions, yaw_chunk = build_action_chunk(deg, clockwise, chunk_frames)
path_ch = os.path.join(subdir, f"_ch{ch}_rand.json")
with open(path_ch, "w") as f:
json.dump(actions, f, indent=2)
if ch == 0:
ctx_pil_0 = load_sample_first_frame(dataset_base, video_name, start_frame, w, h)
if ctx_pil_0 is None:
ctx_pil_0 = sample_random_frame_from_dataset(dataset_base, w, h, seed, metadata_path)
if ctx_pil_0 is not None:
ctx_pil_0 = [ctx_pil_0]
pipe.load_models_to_device(["vae"])
with torch.no_grad():
context_latents = encode_context_frames_per_frame(pipe, ctx_pil_0, pipe.device)
context_actions_t = torch.tensor([identity_rt], dtype=torch.float32)
context_frames_per_chunk.append([ctx_pil_0[0]])
else:
context_frames_per_chunk.append([])
if context_latents is not None and context_actions_t is not None:
frames_ch = run_one_chunk(
pipe, prompt, use_negative_prompt, path_ch,
context_latents=context_latents,
num_context_frames=context_latents.shape[2],
context_actions_t=context_actions_t,
chunk_frames=chunk_frames, h=h, w=w, seed=seed + ch * 7,
sigma_shift=sigma_shift, num_inference_steps=num_inference_steps,
cfg_scale=cfg_scale, inference_noise_level=inference_noise_level,
omit_context_actions=omit_context_actions,
)
else:
frames_ch = run_one_chunk(
pipe, prompt, use_negative_prompt, path_ch,
chunk_frames=chunk_frames, h=h, w=w, seed=seed + ch * 7,
sigma_shift=sigma_shift, num_inference_steps=num_inference_steps, cfg_scale=cfg_scale,
omit_context_actions=omit_context_actions,
)
if ch == 0 and not context_frames_per_chunk[0]:
context_frames_per_chunk[0] = [_frame_to_pil(frames_ch[-1], w, h)]
for y in yaw_chunk:
yaw_history.append(cumulative_yaw_start + y)
cumulative_yaw_start = yaw_history[-1]
chunk_frames_list.append(frames_ch)
# 与 2chunk 一致:仅当有下一 chunk 时用本 chunk 输出做 context(last n 帧 → 逐帧 VAE)
if ch < 3:
n_ctx = min(context_frames, len(frames_ch))
prev_frames = context_frames_for_next_chunk(frames_ch, n_ctx) if n_ctx else [frames_ch[-1]]
prev_pil = [_frame_to_pil(f, w, h) for f in prev_frames]
context_frames_per_chunk.append(list(prev_pil))
pipe.load_models_to_device(["vae"])
with torch.no_grad():
context_latents = encode_context_frames_per_frame(pipe, prev_pil, pipe.device)
num_ctx_tokens = context_latents.shape[2]
context_actions_t = torch.tensor([identity_rt] * num_ctx_tokens, dtype=torch.float32)
print(f"[Loop] next continuation chunk: len(prev_pil)={len(prev_pil)} num_ctx_tokens={num_ctx_tokens} (应等于 ctx)")
if os.path.exists(path_ch):
try:
os.remove(path_ch)
except Exception:
pass
return chunk_frames_list, yaw_history, context_frames_per_chunk
def main():
p = argparse.ArgumentParser(description="Loop: 两栏 [输入轨迹|生成视频]。首帧随机采样训练数据,之后每段 context=上一 chunk last n 帧,context 放右侧")
p.add_argument("--ckpt", required=True, help="Step-15000.safetensors path")
p.add_argument("--dataset_base", default=os.environ.get("DATASET_BASE_PATH", "data/Context-as-Memory-Dataset"))
p.add_argument("--output_dir", required=True)
p.add_argument("--num_samples", type=int, default=4)
p.add_argument("--seed", type=int, default=42)
p.add_argument("--chunk_frames", type=int, default=81)
p.add_argument("--height", type=int, default=352)
p.add_argument("--width", type=int, default=640)
p.add_argument("--sigma_shift", type=float, default=5.0, help="与训练 timestep_shift 一致:pre_qkv noise_5→5, noise_10→10, noise_3→3")
p.add_argument("--num_inference_steps", type=int, default=50, help="与训练 sampling_num_inference_steps 一致(如 50)")
p.add_argument("--cfg_scale", type=float, default=5.0)
p.add_argument("--inference_noise_level", type=float, default=0.0, help="与训练 context_fixed_noise_std 一致(通常 0)")
p.add_argument("--no_omit_context_actions", action="store_true", help="训练使用 context_actions 时加此参数;默认 omit(与 ctx=1 训练一致)")
p.add_argument("--deg_per_chunk", type=float, default=45.0, help="回环每 chunk 转角(左转/右转均用此值,默认 45°)")
p.add_argument("--skip_left_right_2chunk", action="store_true", help="跳过 1左1右 2chunk 回环")
p.add_argument("--skip_left2_right2_4chunk", action="store_true", help="跳过 2左2右 4chunk 回环")
p.add_argument("--only_single_left_45", action="store_true", help="仅跑旋转原子:单 chunk 左转 45°")
p.add_argument("--only_single_right_45", action="store_true", help="仅跑旋转原子:单 chunk 右转 45°")
p.add_argument("--sampling_action_dir", type=str, default=None, help="与采样一致的 action JSON 目录(含 action_rotation_left_45.json/right_45.json);默认用本脚本所在目录")
p.add_argument("--context_image", type=str, default=None, help="泛化检查:用指定图片(如 image.png)作为单帧 context 做旋转原子采样,与训练 action 形式一致")
p.add_argument("--run_legacy_loop", action="store_true", help="跑旧版三组:roundtrip/onedir/replay,与下面 skip 配合")
p.add_argument("--roundtrip_deg", type=float, default=30.0, help="[legacy] 2chunk 折返 顺/逆 角度")
p.add_argument("--onedir_total_deg", type=float, default=60.0, help="[legacy] 4chunk 单向总角度")
p.add_argument("--keep_action_jsons", action="store_true", help="保留临时 action json 便于检查")
p.add_argument("--skip_roundtrip", action="store_true", help="[legacy] 跳过 2chunk 折返")
p.add_argument("--skip_onedir", action="store_true", help="[legacy] 跳过 4chunk 单向 60°")
p.add_argument("--skip_replay_4chunk", action="store_true", help="[legacy] 跳过 4chunk Replay")
p.add_argument("--w_traj", type=int, default=320, help="轨迹图宽度")
p.add_argument("--dataset_metadata_path", type=str, default=None, help="Optional CSV; also used for random first-frame sampling")
p.add_argument("--context_frames", type=int, default=1, help="后续 chunk 的 context 用上一 chunk 的 last n 帧,n=context_frames;--run_ctx5_multi_chunk 时表示 K(K-1 均匀+最后 1 帧)")
p.add_argument("--run_ctx5_multi_chunk", action="store_true", help="跑 ctx5 风格:chunk1 无 context,chunk2 用 K-1 均匀+最后 1 帧;多场景 2chunk 前进/后退/平移/左45右45")
p.add_argument("--run_atomic_translation_4chunk", action="store_true", help="原子操作:仅跑纯平移 4chunk(向前/向后/向左/向右各 4chunk,镜头不旋转)")
p.add_argument("--translation_delta", type=float, default=0.1, help="纯平移每 chunk 位移量(单轴),默认 0.1")
# Memory baselines runtime flags (must match training for multichunk consistency)
p.add_argument("--use_framepack_memory", action="store_true", help="Enable FAR/FramePack-style context reweighting during multichunk inference")
p.add_argument("--context_temporal_decay", type=float, default=1.0, help="FramePack/FAR: per-frame decay for context")
p.add_argument("--context_attention_weight", type=float, default=1.0, help="FramePack/FAR: global scale for all context tokens")
p.add_argument("--use_framepack_length_compress", action="store_true", help="FramePack: compress context length K->K' during multichunk inference")
p.add_argument("--framepack_ratio", type=int, default=2, help="FramePack temporal ratio r")
p.add_argument("--use_spatial_memory", action="store_true", help="Enable spatial memory baseline during multichunk inference")
p.add_argument("--use_spatial_memory_legacy", action="store_true", help="Non-learnable adaptive pool (if no SpatialGridMemory in ckpt, auto-enabled)")
p.add_argument("--spatial_memory_tokens", type=int, default=64, help="Spatial memory baseline: number of pooled spatial tokens")
p.add_argument(
"--spatial_memory_inject_mode",
type=str,
default=None,
choices=("concat_text", "cross_attn_readout", "none"),
help="Spatial memory inject mode; must match training",
)
p.add_argument("--num_gpus", type=int, default=1, help="总 GPU 数,与 --rank 配合做多卡并行:每卡只推理 rank, rank+num_gpus, ... 的样本")
p.add_argument("--rank", type=int, default=0, help="当前进程的 rank(0 到 num_gpus-1),由启动脚本设 CUDA_VISIBLE_DEVICES=rank 并传 --rank")
p.add_argument(
"--multi_ctx_4chunk_all_history",
action="store_true",
help="4chunk 回环:续段 ctx 时,除最后 1 帧固定为上一 chunk 的最后一帧,其余 (ctx-1) 帧从所有已生成 chunk 的历史帧中均匀抽取,用于多 chunk 记忆机制实验",
)
p.add_argument(
"--multi_ctx_4chunk_fov_history",
action="store_true",
help="4chunk 回环:续段 ctx 时,除上一 chunk 最后一帧外,其余帧按下一 chunk 中点 world-yaw 从已生成历史帧中检索(仅改 context 帧,context_actions 仍保持 identity)",
)
p.add_argument(
"--multi_ctx_4chunk_fov_last_target",
action="store_true",
help="配合 --multi_ctx_4chunk_fov_history:仅第 4 个 chunk 的检索 target 改用该 chunk 末尾 world-yaw,便于回到起点时检索 chunk1 开头帧",
)
p.add_argument(
"--multi_ctx_4chunk_fov_context_rt",
action="store_true",
help="配合 --multi_ctx_4chunk_fov_history:为检索出的 context 帧注入相对下一 chunk 起始 world-yaw 的 RT,而不是 identity",
)
# Camera encoder 与注入节点:须与训练一致(见 INFERENCE_ALIGNMENT.md)
p.add_argument("--camera_inject_mode", type=str, default=None, help="与训练一致:post|pre_norm|pre_qkv|pre_qkv_post;不设则从 ckpt 路径推断(如含 pre_qkv_post)")
p.add_argument("--camera_encoder_shallow", action="store_true", default=None, help="与训练一致:单层 Linear(12,D)。不设则从 ckpt 自动推断")
p.add_argument("--no_camera_encoder_shallow", action="store_true", help="明确不用 shallow(3 层 MLP)")
p.add_argument("--camera_encoder_separate_t_r", action="store_true", default=None, help="与训练一致:t/R 分路编码。不设则从 ckpt 自动推断")
p.add_argument("--no_camera_encoder_separate_t_r", action="store_true", help="明确不用 separate_t_r")
p.add_argument("--no_camera_encoder_zero_init", action="store_true", help="与训练一致:encoder scale 非 0 初始化时加此参数")
p.add_argument("--camera_encoder_explicit_yaw", action="store_true", help="与训练一致:显式 yaw 分支(dir_abl_A 等)")
p.add_argument("--camera_encoder_sincos_yaw", action="store_true", help="与训练一致:sin/cos yaw 分支")
p.add_argument("--add_camera_outside_gate", action="store_true", help="与训练一致:camera 加在 gate 外(abl 实验)")
p.add_argument(
"--base_model",
type=str,
default=None,
help="Wan2.1 base model dir; default: $WAN_BASE_MODEL",
)
args = p.parse_args()
# 原子平移评估用独立前缀,与回环 [Loop] 区分
_log_prefix = "[atomic_translation_4chunk]" if getattr(args, "run_atomic_translation_4chunk", False) else "[Loop]"
# 与训练 setting 对齐:ctx=1 时训练用 omit_context_actions=True
args.omit_context_actions = not getattr(args, "no_omit_context_actions", False)
# context_k* / ctx_*: infer context frames from released HF folders and legacy names.
ckpt_lower = (args.ckpt or "").lower()
if "ctx_20" in ckpt_lower or "context_k20" in ckpt_lower or "ctx20" in ckpt_lower:
if getattr(args, "context_frames", 1) < 20:
args.context_frames = 20
print(f"{_log_prefix} 从 ckpt 路径识别 context_k20,将 context_frames 设为 20(续段 20 帧逐帧 VAE)", flush=True)
elif "ctx_5" in ckpt_lower or "context_k5" in ckpt_lower or "ctx5" in ckpt_lower:
if getattr(args, "context_frames", 1) < 5:
args.context_frames = 5
print(f"{_log_prefix} 从 ckpt 路径识别 context_k5,将 context_frames 设为 5(续段 5 帧)", flush=True)
# 兼容:run_ctx5_multi_chunk 时若尚未被上面推断,也设为 5
if getattr(args, "run_ctx5_multi_chunk", False) and getattr(args, "context_frames", 1) == 1:
args.context_frames = 5
print(f"{_log_prefix} run_ctx5_multi_chunk 已启用,将 context_frames 设为 5(续段 ctx=5)", flush=True)
print(f"{_log_prefix} context_frames={args.context_frames}(续段用上一 chunk 的 last n 帧;首 chunk 始终 1 帧)", flush=True)
# camera_inject_mode:未显式指定时从 ckpt 路径推断(与训练脚本命名一致)
camera_inject_mode = (getattr(args, "camera_inject_mode", None) or "").strip()
if not camera_inject_mode:
env_cam = (os.environ.get("CAMERA_INJECT_MODE") or "").strip().lower()
if env_cam in ("pre_qkv_post", "pre_qkv", "pre_norm", "post"):
camera_inject_mode = env_cam
print(f"{_log_prefix} 从环境 CAMERA_INJECT_MODE 使用 camera_inject_mode={camera_inject_mode}", flush=True)
if not camera_inject_mode:
ckpt_lower = (args.ckpt or "").lower()
for mode in ("pre_qkv_post", "pre_qkv", "pre_norm", "post"):
if mode.replace("_", "") in ckpt_lower or mode in ckpt_lower:
camera_inject_mode = mode
print(f"{_log_prefix} 从 ckpt 路径推断 camera_inject_mode={camera_inject_mode}", flush=True)
break
if not camera_inject_mode:
# 与 memory_baselines_basic 默认训练脚本一致(旧默认 post 易与 pre_qkv 训练 ckpt 错位)
camera_inject_mode = "pre_qkv"
print(f"{_log_prefix} camera_inject_mode 未推断,使用默认 pre_qkv(与 train/memory_baselines_basic 对齐)", flush=True)
# encoder 结构、SSM:均由 load_pipeline_and_ckpt 按 ckpt 自动推断并注入,无需额外参数
load_kw = dict(
action_inject_after_spatial_attn=True,
add_action_attn=True,
action_use_temporal_attention=True,
camera_inject_mode=camera_inject_mode,
add_camera_outside_gate=getattr(args, "add_camera_outside_gate", False),
no_camera_encoder_zero_init=getattr(args, "no_camera_encoder_zero_init", False),
camera_encoder_explicit_yaw=getattr(args, "camera_encoder_explicit_yaw", False),
camera_encoder_sincos_yaw=getattr(args, "camera_encoder_sincos_yaw", False),
)
if getattr(args, "camera_encoder_shallow", False):
load_kw["camera_encoder_shallow"] = True
if getattr(args, "no_camera_encoder_shallow", False):
load_kw["camera_encoder_shallow"] = False
if getattr(args, "camera_encoder_separate_t_r", False):
load_kw["camera_encoder_separate_t_r"] = True
if getattr(args, "no_camera_encoder_separate_t_r", False):
load_kw["camera_encoder_separate_t_r"] = False
if args.num_gpus > 1:
cuda_vis = os.environ.get("CUDA_VISIBLE_DEVICES", "not set")
print(f"{_log_prefix} rank={args.rank} CUDA_VISIBLE_DEVICES={cuda_vis}", flush=True)
base_model = (
getattr(args, "base_model", None)
or os.environ.get("WAN_BASE_MODEL")
)
if not base_model:
raise ValueError("Set --base_model or WAN_BASE_MODEL to the Wan2.1 base model directory.")
dit_path = os.path.join(base_model, "diffusion_pytorch_model.safetensors")
text_path = os.path.join(base_model, "models_t5_umt5-xxl-enc-bf16.pth")
vae_path = os.path.join(base_model, "Wan2.1_VAE.pth")
for _p in (dit_path, text_path, vae_path):
if not os.path.isfile(_p):
raise FileNotFoundError(
f"Missing Wan2.1 base weight: {_p} (set --base_model or WAN_BASE_MODEL to override)"
)
pipe = irc.load_pipeline_and_ckpt(
args.ckpt,
dit_path,
text_path,
vae_path,
**load_kw,
)
# Runtime memory flags (must align with the checkpoint training flags)
pipe.use_framepack_memory = bool(getattr(args, "use_framepack_memory", False))
pipe.context_temporal_decay = float(getattr(args, "context_temporal_decay", 1.0) or 1.0)
pipe.context_attention_weight = float(getattr(args, "context_attention_weight", 1.0) or 1.0)
pipe.use_framepack_length_compress = bool(getattr(args, "use_framepack_length_compress", False))
pipe.framepack_ratio = int(getattr(args, "framepack_ratio", 2) or 2)
pipe.use_spatial_memory = bool(getattr(args, "use_spatial_memory", False))
pipe.spatial_memory_tokens = int(getattr(args, "spatial_memory_tokens", 64) or 64)
if getattr(args, "spatial_memory_inject_mode", None):
pipe.spatial_memory_inject_mode = str(getattr(args, "spatial_memory_inject_mode"))
pipe.use_spatial_memory_legacy = bool(getattr(args, "use_spatial_memory_legacy", False))
# Ckpt may omit SpatialGridMemory weights: fall back to legacy adaptive pool for correct shapes
if pipe.use_spatial_memory and not pipe.use_spatial_memory_legacy and getattr(pipe, "spatial_memory_module", None) is None:
pipe.use_spatial_memory_legacy = True
print(f"{_log_prefix} use_spatial_memory but no spatial_memory_module in ckpt -> use_spatial_memory_legacy=True", flush=True)
if args.num_gpus > 1 and torch.cuda.is_available():
print(f"{_log_prefix} rank={args.rank} device_count={torch.cuda.device_count()} current_device={torch.cuda.current_device()}", flush=True)
if args.dataset_metadata_path and os.path.isfile(args.dataset_metadata_path):
import csv
candidates = []
with open(args.dataset_metadata_path, "r", encoding="utf-8") as f:
reader = csv.DictReader(f)
for row in reader:
vn = row.get("video_name", "").strip()
sf = row.get("start_frame", "")
if vn and str(sf).strip():
try:
candidates.append((vn, int(sf)))
except ValueError:
continue
rng = random.Random(args.seed)
if len(candidates) <= args.num_samples:
samples = candidates
else:
samples = [candidates[i] for i in rng.sample(range(len(candidates)), args.num_samples)]
else:
samples = irc.sample_trajectory_samples_from_dataset(
args.dataset_base,
num_samples=args.num_samples,
num_frames=args.chunk_frames,
seed=args.seed,
)
if not samples:
print(f"{_log_prefix} No samples from dataset.")
return
# 多卡:每卡只跑自己 rank 的样本(样本顺序一致,按 idx % num_gpus == rank 划分)
if args.num_gpus > 1:
samples = [s for i, s in enumerate(samples) if i % args.num_gpus == args.rank]
print(f"{_log_prefix} num_gpus={args.num_gpus} rank={args.rank} -> 本进程跑 {len(samples)} 个样本")
if not samples:
print(f"{_log_prefix} 本 rank 无样本,退出")
return
os.makedirs(args.output_dir, exist_ok=True)
h, w = args.height, args.width
# 日志同时写入 output_dir 下的文件
log_path = os.path.join(args.output_dir, f"loop_run_{datetime.now().strftime('%Y%m%d_%H%M%S')}.log")
_log_file = open(log_path, "w", encoding="utf-8")
_stdout_orig = sys.stdout
class _Tee:
def __init__(self, stream, f):
self._stream = stream
self._file = f
def write(self, data):
self._stream.write(data)
self._stream.flush()
if self._file and not self._file.closed:
self._file.write(data)
self._file.flush()
def flush(self):
self._stream.flush()
if self._file and not self._file.closed:
self._file.flush()
sys.stdout = _Tee(_stdout_orig, _log_file)
try:
print(f"{_log_prefix} Log file: {log_path}")
if getattr(args, "run_atomic_translation_4chunk", False):
print(f"{_log_prefix} 模式: 仅原子操作纯平移 4chunk (forward/backward/left/right),无旋转", flush=True)
elif getattr(args, "run_ctx5_multi_chunk", False):
print(f"{_log_prefix} 模式: ctx5 多 chunk 2chunk 多场景 (含旋转)", flush=True)
else:
print(f"{_log_prefix} 模式: 回环 1左1右 + 2左2右 旋转 (若只要纯平移请加 --run_atomic_translation_4chunk)", flush=True)
for idx, (video_name, start_frame) in enumerate(samples):
subdir = os.path.join(args.output_dir, f"loop_{video_name}_start{start_frame}")
os.makedirs(subdir, exist_ok=True)
# 同一样本 (video_name, start_frame) 使用确定性 seed,与 loop_traj_fov_gen / evals_ep0 等调用一致,避免偏移
sample_seed = seed_for_sample(args.seed, video_name, start_frame)
print(f"{_log_prefix} [{idx+1}/{len(samples)}] {video_name} start{start_frame} sigma_shift={args.sigma_shift} sample_seed={sample_seed}")
def save_composite(name, chunk_frames_list, yaw_history, context_frames_per_chunk):
# 与训练对齐:续段首帧=上一段末帧,拼接时去掉续段第 0 帧,避免重复/跳帧(chunk0 保留 81,chunk1+ 各 80)
if len(chunk_frames_list) > 1:
trim_chunks, trim_yaw = trim_continuation_first_frame(chunk_frames_list, yaw_history, args.chunk_frames)
else:
trim_chunks, trim_yaw = chunk_frames_list, yaw_history
right_frames, yaw_ext = build_right_panel_and_yaw(
trim_chunks, trim_yaw, context_frames_per_chunk,
chunk_frames=args.chunk_frames, w=w, h=h,
)
traj_frames = draw_trajectory_frames(yaw_ext, args.w_traj, h, total_frames=len(right_frames))
comp = composite_two_panel(traj_frames, right_frames, w_traj=args.w_traj, h_traj=h, w_gen=w, h_gen=h)
save_video(comp, os.path.join(subdir, f"{name}_traj_gen.mp4"), fps=15, quality=5)
gen_only = [f for ch in trim_chunks for f in ch]
save_video(gen_only, os.path.join(subdir, f"{name}_gen_only.mp4"), fps=15, quality=5)
meta = getattr(args, "dataset_metadata_path", None)
try:
# 原子操作:仅纯平移 4chunk(无旋转),与回环左/右转彻底互斥
if getattr(args, "run_atomic_translation_4chunk", False):
delta = getattr(args, "translation_delta", 0.1)
for direction in ("forward", "backward", "left", "right"):
chunks, yaw, ctx_per_chunk = run_translation_4chunk(
pipe, args.dataset_base, args.output_dir, video_name, start_frame,
direction=direction, translation_delta=delta, chunk_frames=args.chunk_frames,
context_frames=args.context_frames, h=h, w=w,
sigma_shift=args.sigma_shift, num_inference_steps=args.num_inference_steps,
cfg_scale=args.cfg_scale, seed=sample_seed, inference_noise_level=args.inference_noise_level,
metadata_path=meta, omit_context_actions=args.omit_context_actions,
)
save_composite(f"{direction}_4chunk", chunks, yaw, ctx_per_chunk)
# 旋转原子:仅单 chunk 左转 45° 或 右转 45°
elif getattr(args, "only_single_left_45", False):
chunks, yaw, ctx_per_chunk = run_single_chunk_rotation(
pipe, args.dataset_base, args.output_dir, video_name, start_frame,
deg=args.deg_per_chunk, clockwise=False, chunk_frames=args.chunk_frames,
h=h, w=w, sigma_shift=args.sigma_shift, num_inference_steps=args.num_inference_steps,
cfg_scale=args.cfg_scale, seed=sample_seed, inference_noise_level=args.inference_noise_level,
metadata_path=meta, keep_action_jsons=getattr(args, "keep_action_jsons", False),
sampling_action_dir=getattr(args, "sampling_action_dir", None),
omit_context_actions=args.omit_context_actions,
context_image_path=getattr(args, "context_image", None),
)
save_composite("single_left_45", chunks, yaw, ctx_per_chunk)
elif getattr(args, "only_single_right_45", False):
chunks, yaw, ctx_per_chunk = run_single_chunk_rotation(
pipe, args.dataset_base, args.output_dir, video_name, start_frame,
deg=args.deg_per_chunk, clockwise=True, chunk_frames=args.chunk_frames,
h=h, w=w, sigma_shift=args.sigma_shift, num_inference_steps=args.num_inference_steps,
cfg_scale=args.cfg_scale, seed=sample_seed, inference_noise_level=args.inference_noise_level,
metadata_path=meta, keep_action_jsons=getattr(args, "keep_action_jsons", False),
sampling_action_dir=getattr(args, "sampling_action_dir", None),
omit_context_actions=args.omit_context_actions,
context_image_path=getattr(args, "context_image", None),
)
save_composite("single_right_45", chunks, yaw, ctx_per_chunk)
# 默认:回环两场景 1左1右(45°)、2左2右(每chunk 45°)。--run_legacy_loop 时跑旧三组
elif not getattr(args, "run_legacy_loop", False):
if not getattr(args, "skip_left_right_2chunk", False):
chunks, yaw, ctx_per_chunk = run_left_right_2chunk(
pipe, args.dataset_base, args.output_dir, video_name, start_frame,
deg=args.deg_per_chunk, chunk_frames=args.chunk_frames, context_frames=args.context_frames,
h=h, w=w, sigma_shift=args.sigma_shift, num_inference_steps=args.num_inference_steps,
cfg_scale=args.cfg_scale, seed=sample_seed, inference_noise_level=args.inference_noise_level,
metadata_path=meta, keep_action_jsons=getattr(args, "keep_action_jsons", False),
sampling_action_dir=getattr(args, "sampling_action_dir", None),
omit_context_actions=args.omit_context_actions,
)
save_composite("left_right_2chunk", chunks, yaw, ctx_per_chunk)
if not getattr(args, "skip_left2_right2_4chunk", False):
chunks, yaw, ctx_per_chunk = run_left2_right2_4chunk(
pipe, args.dataset_base, args.output_dir, video_name, start_frame,
deg_per_chunk=args.deg_per_chunk, chunk_frames=args.chunk_frames, context_frames=args.context_frames,
h=h, w=w, sigma_shift=args.sigma_shift, num_inference_steps=args.num_inference_steps,
cfg_scale=args.cfg_scale, seed=sample_seed + 100, inference_noise_level=args.inference_noise_level,
metadata_path=meta, sampling_action_dir=getattr(args, "sampling_action_dir", None),
omit_context_actions=args.omit_context_actions,
multi_ctx_all_history=getattr(args, "multi_ctx_4chunk_all_history", False),
fov_history_context=getattr(args, "multi_ctx_4chunk_fov_history", False),
fov_last_target=getattr(args, "multi_ctx_4chunk_fov_last_target", False),
fov_context_rt=getattr(args, "multi_ctx_4chunk_fov_context_rt", False),
)
save_composite("left2_right2_4chunk", chunks, yaw, ctx_per_chunk)
else:
if not args.skip_roundtrip:
chunks, yaw, ctx_per_chunk = run_roundtrip_2chunk(
pipe, args.dataset_base, args.output_dir, video_name, start_frame,
deg=args.roundtrip_deg, chunk_frames=args.chunk_frames, context_frames=args.context_frames,
h=h, w=w, sigma_shift=args.sigma_shift, num_inference_steps=args.num_inference_steps,
cfg_scale=args.cfg_scale, seed=sample_seed, inference_noise_level=args.inference_noise_level,
metadata_path=meta, keep_action_jsons=getattr(args, "keep_action_jsons", False),
omit_context_actions=args.omit_context_actions,
)
save_composite("roundtrip_2chunk", chunks, yaw, ctx_per_chunk)
if not args.skip_onedir:
deg_per = args.onedir_total_deg / 4.0
chunks, yaw, ctx_per_chunk = run_one_direction_4chunk(
pipe, args.dataset_base, args.output_dir, video_name, start_frame,
deg_per_chunk=deg_per, clockwise=True, chunk_frames=args.chunk_frames,
context_frames=args.context_frames, h=h, w=w,
sigma_shift=args.sigma_shift, num_inference_steps=args.num_inference_steps,
cfg_scale=args.cfg_scale, seed=sample_seed + 100, inference_noise_level=args.inference_noise_level,
metadata_path=meta, omit_context_actions=args.omit_context_actions,
)
save_composite("onedir_4chunk_60", chunks, yaw, ctx_per_chunk)
if not args.skip_replay_4chunk:
result = run_replay_4chunk(
pipe, args.dataset_base, args.output_dir, video_name, start_frame,
chunk_frames=args.chunk_frames, context_frames=args.context_frames,
h=h, w=w, sigma_shift=args.sigma_shift, num_inference_steps=args.num_inference_steps,
cfg_scale=args.cfg_scale, seed=sample_seed + 200, inference_noise_level=args.inference_noise_level,
omit_context_actions=args.omit_context_actions,
)
if result is not None:
chunks, yaw, ctx_per_chunk = result
save_composite("replay_4chunk", chunks, yaw, ctx_per_chunk)
else:
print(f"[Loop] Skip replay_4chunk for {video_name} start{start_frame} (need 4*81 frames)")
except Exception as e:
print(f"[Loop] ERROR {video_name} start{start_frame}: {e}")
import traceback
traceback.print_exc()
if getattr(args, "only_single_left_45", False):
mode = "旋转原子 单chunk 左转45°"
elif getattr(args, "only_single_right_45", False):
mode = "旋转原子 单chunk 右转45°"
elif getattr(args, "run_legacy_loop", False):
mode = "legacy"
else:
mode = "回环 1左1右 + 2左2右 (每chunk 45°)"
print(f"{_log_prefix} Done. Mode: {mode}. Outputs under {args.output_dir}")
finally:
sys.stdout = _stdout_orig
_log_file.close()
if __name__ == "__main__":
main()
|