()
| 38 | |
| 39 | |
| 40 | def get_preprocess_parser(): |
| 41 | parser = argparse.ArgumentParser(description="LTX-2.3 IC-LoRA control-video preprocessing (pose / canny / depth) for v2av.") |
| 42 | |
| 43 | parser.add_argument( |
| 44 | "--mode", |
| 45 | type=str, |
| 46 | default="pose", |
| 47 | choices=["pose", "canny", "depth"], |
| 48 | help="pose: DWPose skeleton (needs --ckpt_path). canny: OpenCV Canny edges (Union-Control). depth: MiDaS-small monocular depth via torch.hub (Union-Control; first run downloads weights).", |
| 49 | ) |
| 50 | |
| 51 | parser.add_argument( |
| 52 | "--ckpt_path", |
| 53 | type=str, |
| 54 | default=None, |
| 55 | help="Required when --mode pose: DWPose ONNX root with det/yolox_l.onnx and pose2d/dw-ll_ucoco_384.onnx (see yzd-v/DWPose). Ignored for canny/depth.", |
| 56 | ) |
| 57 | |
| 58 | parser.add_argument("--video_path", type=str, default=None, help="The path to the driving video.") |
| 59 | parser.add_argument( |
| 60 | "--refer_path", |
| 61 | type=str, |
| 62 | default=None, |
| 63 | help="Optional path to the character/reference image (the human appearance the " |
| 64 | "downstream LTX-2.3 v2av task will animate). When provided, the skeleton canvas " |
| 65 | "aspect ratio is taken from this image; source video " |
| 66 | "frames are letterboxed into the canvas before DWPose runs. This guarantees a " |
| 67 | "portrait image yields a portrait output, etc. When omitted, the canvas falls " |
| 68 | "back to the source video aspect ratio.", |
| 69 | ) |
| 70 | parser.add_argument( |
| 71 | "--save_path", |
| 72 | type=str, |
| 73 | default=None, |
| 74 | help="Output control video (.mp4). If a directory is given, writes pose_skeleton.mp4 / canny_control.mp4 / depth_control.mp4 depending on --mode.", |
| 75 | ) |
| 76 | |
| 77 | parser.add_argument( |
| 78 | "--resolution_area", |
| 79 | type=int, |
| 80 | nargs=2, |
| 81 | default=[1280, 720], |
| 82 | help="The target resolution area for the skeleton video, specified as " |
| 83 | "[width, height]. The video (or, if --refer_path is given, the image) is " |
| 84 | "resized to have a total area equivalent to width * height while preserving " |
| 85 | "its original aspect ratio. Both dimensions are snapped to a multiple of 32 " |
| 86 | "for LTX-2.3.", |
| 87 | ) |
| 88 | parser.add_argument( |
| 89 | "--fps", |
| 90 | type=int, |
| 91 | default=24, |
| 92 | help="Target FPS for the output skeleton video. Set to -1 to keep the source FPS. LTX-2.3 IC-LoRA is typically trained at 24 FPS.", |
| 93 | ) |
| 94 | parser.add_argument( |
| 95 | "--num_frames", |
| 96 | type=int, |
| 97 | default=-1, |
no outgoing calls
no test coverage detected