diff --git a/benchmarks/edit/build_six_method_manifest.py b/benchmarks/edit/build_six_method_manifest.py new file mode 100644 index 0000000000000000000000000000000000000000..624b99c270ab3114fcd2b43648bd7294717bb22a --- /dev/null +++ b/benchmarks/edit/build_six_method_manifest.py @@ -0,0 +1,82 @@ +#!/usr/bin/env python3 +"""Build manifests for the six val20/val100 edit-result folders.""" + +from __future__ import annotations + +import argparse +import json +from pathlib import Path + + +METHODS = ("wan_only", "ditto_global", "full", "text", "vace_hint", "vace_context") + + +def read_jsonl(path: Path) -> list[dict]: + rows: list[dict] = [] + with path.open("r", encoding="utf-8") as handle: + for line in handle: + text = line.strip() + if text: + rows.append(json.loads(text)) + return rows + + +def build_split(repo_root: Path, split: str, samples_path: Path, outputs_root: Path, output_path: Path) -> None: + samples = read_jsonl(samples_path) + output_path.parent.mkdir(parents=True, exist_ok=True) + count = 0 + with output_path.open("w", encoding="utf-8") as handle: + for sample in samples: + sample_id = str(sample["id"]) + source_video = Path(str(sample["control_video"])) + if not source_video.is_absolute(): + source_video = repo_root / source_video + for method in METHODS: + edited_video = outputs_root / method / f"{sample_id}.mp4" + row = { + "split": split, + "sample_id": sample_id, + "method": method, + "instruction": sample["prompt"], + "source_video": str(source_video), + "edited_video": str(edited_video), + "source_prompt": sample.get("source_prompt", ""), + "target_prompt": sample.get("target_prompt", sample["prompt"]), + } + handle.write(json.dumps(row, ensure_ascii=False) + "\n") + count += 1 + print(f"{split}: wrote {count} rows -> {output_path}") + + +def main() -> None: + parser = argparse.ArgumentParser() + parser.add_argument("--repo-root", type=Path, default=Path.cwd()) + parser.add_argument("--output-dir", type=Path, default=Path("out/edit_model_face_stage1/traditional_eval_manifests")) + args = parser.parse_args() + + repo_root = args.repo_root.resolve() + base = repo_root / "out/edit_model_face_stage1" + output_dir = args.output_dir if args.output_dir.is_absolute() else repo_root / args.output_dir + + jobs = [ + ( + "val20", + base / "eval_samples/val_20.jsonl", + base / "eval_outputs", + output_dir / "val20.jsonl", + ), + ( + "val100", + base / "eval_samples/val_100.jsonl", + base / "eval_outputs_val100", + output_dir / "val100.jsonl", + ), + ] + for split, samples_path, outputs_root, output_path in jobs: + if not samples_path.exists(): + raise FileNotFoundError(samples_path) + build_split(repo_root, split, samples_path, outputs_root, output_path) + + +if __name__ == "__main__": + main() diff --git a/benchmarks/edit/code/IVEBench/metrics/__pycache__/ivebench.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/__pycache__/ivebench.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..2aea6e1190c4d569a37a2dc9086f83c1c93b9c5f Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/__pycache__/ivebench.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/__pycache__/ivebench_utils.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/__pycache__/ivebench_utils.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..c977532e53e3fc7f10a54d0747186b8ad14d8ed3 Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/__pycache__/ivebench_utils.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/__pycache__/vebench.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/__pycache__/vebench.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..26bc23eaa5cd59fd9d9b5218e7b82ea0fd3c3209 Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/__pycache__/vebench.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/compliance/__pycache__/__init__.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/compliance/__pycache__/__init__.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..e5fc353369fbd0f03e418e0d802c1c0d25ac04aa Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/compliance/__pycache__/__init__.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/compliance/__pycache__/instruction_satisfaction.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/compliance/__pycache__/instruction_satisfaction.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..36b996976e97817e2bb97d4deb8b927d2118bd38 Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/compliance/__pycache__/instruction_satisfaction.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/compliance/__pycache__/overall_semantic_consistency.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/compliance/__pycache__/overall_semantic_consistency.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..dcc66aa683ba99c4c2c43c0e40495fb3ac030adf Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/compliance/__pycache__/overall_semantic_consistency.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/compliance/__pycache__/phrase_semantic_consistency.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/compliance/__pycache__/phrase_semantic_consistency.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..fb593cee76c8158b1b5d6c2d408c33fabdbbf95f Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/compliance/__pycache__/phrase_semantic_consistency.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/compliance/__pycache__/quantity_accuracy.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/compliance/__pycache__/quantity_accuracy.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..3884f1cc5bce12fae9512cd9d31db2dc134ee0a8 Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/compliance/__pycache__/quantity_accuracy.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/compliance/groundingdino/__pycache__/__init__.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/compliance/groundingdino/__pycache__/__init__.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..8b89212747ec49c412b5b99596016d7f69c9c4e3 Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/compliance/groundingdino/__pycache__/__init__.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/compliance/groundingdino/datasets/__pycache__/__init__.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/compliance/groundingdino/datasets/__pycache__/__init__.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..586fee799778009019a9335ac73ea4f299a72f59 Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/compliance/groundingdino/datasets/__pycache__/__init__.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/compliance/groundingdino/datasets/__pycache__/transforms.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/compliance/groundingdino/datasets/__pycache__/transforms.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..84ba9347f6a768aed2671be17898e120a0169cb8 Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/compliance/groundingdino/datasets/__pycache__/transforms.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/compliance/groundingdino/models/GroundingDINO/__pycache__/__init__.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/compliance/groundingdino/models/GroundingDINO/__pycache__/__init__.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..9b0d61a4779517edd524316e39f71bf748fa2f11 Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/compliance/groundingdino/models/GroundingDINO/__pycache__/__init__.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/compliance/groundingdino/models/GroundingDINO/__pycache__/bertwarper.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/compliance/groundingdino/models/GroundingDINO/__pycache__/bertwarper.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..9f89492c375363da1806e847a0479d894a2d8f0d Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/compliance/groundingdino/models/GroundingDINO/__pycache__/bertwarper.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/compliance/groundingdino/models/GroundingDINO/__pycache__/fuse_modules.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/compliance/groundingdino/models/GroundingDINO/__pycache__/fuse_modules.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..b2219bf8b3ed6c537fe83df72e20b5cf726df089 Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/compliance/groundingdino/models/GroundingDINO/__pycache__/fuse_modules.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/compliance/groundingdino/models/GroundingDINO/__pycache__/groundingdino.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/compliance/groundingdino/models/GroundingDINO/__pycache__/groundingdino.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..2c5b5ae1bc8fc8d7ecb1f8c54ad2d6bfcfaaa1b9 Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/compliance/groundingdino/models/GroundingDINO/__pycache__/groundingdino.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/compliance/groundingdino/models/GroundingDINO/__pycache__/ms_deform_attn.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/compliance/groundingdino/models/GroundingDINO/__pycache__/ms_deform_attn.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..e2d45794a0019c966b72b5bc73858d2ed305c246 Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/compliance/groundingdino/models/GroundingDINO/__pycache__/ms_deform_attn.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/compliance/groundingdino/models/GroundingDINO/__pycache__/transformer.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/compliance/groundingdino/models/GroundingDINO/__pycache__/transformer.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..cb7a4fccb26da3a6d6c5e9a0a906f3f72b46f7e7 Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/compliance/groundingdino/models/GroundingDINO/__pycache__/transformer.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/compliance/groundingdino/models/GroundingDINO/__pycache__/transformer_vanilla.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/compliance/groundingdino/models/GroundingDINO/__pycache__/transformer_vanilla.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..dc3d4a59f207009ee3d9d3bd180c0871e53a9a9b Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/compliance/groundingdino/models/GroundingDINO/__pycache__/transformer_vanilla.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/compliance/groundingdino/models/GroundingDINO/__pycache__/utils.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/compliance/groundingdino/models/GroundingDINO/__pycache__/utils.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..bc107061e236e51442349a94f13f1a872b5e1f92 Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/compliance/groundingdino/models/GroundingDINO/__pycache__/utils.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/compliance/groundingdino/models/GroundingDINO/backbone/__pycache__/__init__.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/compliance/groundingdino/models/GroundingDINO/backbone/__pycache__/__init__.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..ee3a94b1ac00796340ac3087e780f38fa4ca4e5f Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/compliance/groundingdino/models/GroundingDINO/backbone/__pycache__/__init__.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/compliance/groundingdino/models/GroundingDINO/backbone/__pycache__/backbone.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/compliance/groundingdino/models/GroundingDINO/backbone/__pycache__/backbone.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..5281135d7ac55c4b6f9611c7b5fb64d1f5795038 Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/compliance/groundingdino/models/GroundingDINO/backbone/__pycache__/backbone.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/compliance/groundingdino/models/GroundingDINO/backbone/__pycache__/position_encoding.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/compliance/groundingdino/models/GroundingDINO/backbone/__pycache__/position_encoding.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..f589021b48b850c1b27bcd5ad3238bd33f72284d Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/compliance/groundingdino/models/GroundingDINO/backbone/__pycache__/position_encoding.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/compliance/groundingdino/models/GroundingDINO/backbone/__pycache__/swin_transformer.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/compliance/groundingdino/models/GroundingDINO/backbone/__pycache__/swin_transformer.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..d37ef484887095f9c0d6ffc08ffd722624d92f6d Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/compliance/groundingdino/models/GroundingDINO/backbone/__pycache__/swin_transformer.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/compliance/groundingdino/models/__pycache__/__init__.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/compliance/groundingdino/models/__pycache__/__init__.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..86b6f2c4a10334ef5a6f81dc267e7231bf646eb7 Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/compliance/groundingdino/models/__pycache__/__init__.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/compliance/groundingdino/models/__pycache__/registry.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/compliance/groundingdino/models/__pycache__/registry.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..8897f529e86f189417807b91d5699339f72d0cea Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/compliance/groundingdino/models/__pycache__/registry.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/compliance/groundingdino/util/__pycache__/__init__.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/compliance/groundingdino/util/__pycache__/__init__.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..c0d94a9049104f5a4dbcff62de0b00bf6cd23b9d Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/compliance/groundingdino/util/__pycache__/__init__.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/compliance/groundingdino/util/__pycache__/box_ops.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/compliance/groundingdino/util/__pycache__/box_ops.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..684d15085c2143a572eb53fa6c5fcaae01820a0d Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/compliance/groundingdino/util/__pycache__/box_ops.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/compliance/groundingdino/util/__pycache__/get_tokenlizer.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/compliance/groundingdino/util/__pycache__/get_tokenlizer.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..7383c848522f42f2054247ab9814aaa92812290f Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/compliance/groundingdino/util/__pycache__/get_tokenlizer.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/compliance/groundingdino/util/__pycache__/misc.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/compliance/groundingdino/util/__pycache__/misc.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..17c201a1d3cd2f5dd845f3988528977e9507dd25 Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/compliance/groundingdino/util/__pycache__/misc.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/compliance/groundingdino/util/__pycache__/slconfig.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/compliance/groundingdino/util/__pycache__/slconfig.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..e06ccb9f03a7bee3ff447fe7ba07b9ed1f6384d5 Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/compliance/groundingdino/util/__pycache__/slconfig.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/compliance/groundingdino/util/__pycache__/utils.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/compliance/groundingdino/util/__pycache__/utils.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..1fb660a98beecf0a6f54ef00c908bcba10ca509a Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/compliance/groundingdino/util/__pycache__/utils.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/compliance/groundingdino/util/__pycache__/visualizer.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/compliance/groundingdino/util/__pycache__/visualizer.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..143822976fa94586e8f58fd5726148226519db63 Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/compliance/groundingdino/util/__pycache__/visualizer.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/compliance/groundingdino/util/__pycache__/vl_utils.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/compliance/groundingdino/util/__pycache__/vl_utils.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..fa6f983437f4fc3bc73850e80d6ebd5c8f05a7f6 Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/compliance/groundingdino/util/__pycache__/vl_utils.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/compliance/qwen_vl_utils/__pycache__/__init__.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/compliance/qwen_vl_utils/__pycache__/__init__.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..384cff873e2db5ecc1e301e007f35f17855902f9 Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/compliance/qwen_vl_utils/__pycache__/__init__.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/compliance/qwen_vl_utils/__pycache__/vision_process.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/compliance/qwen_vl_utils/__pycache__/vision_process.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..09ffd586c260545ab31f03742c499af7b159217c Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/compliance/qwen_vl_utils/__pycache__/vision_process.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/compliance/videoclipxl_utils/__pycache__/__init__.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/compliance/videoclipxl_utils/__pycache__/__init__.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..dce961a6583f9380aff605f37a291056015d53ba Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/compliance/videoclipxl_utils/__pycache__/__init__.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/compliance/videoclipxl_utils/__pycache__/modeling.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/compliance/videoclipxl_utils/__pycache__/modeling.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..4dabcad277eed60bbbd205df739520a4b34b57d5 Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/compliance/videoclipxl_utils/__pycache__/modeling.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/compliance/videoclipxl_utils/text_encoder/__pycache__/__init__.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/compliance/videoclipxl_utils/text_encoder/__pycache__/__init__.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..c88adc6e56eeebaac96d36d2c28f562ee98155a2 Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/compliance/videoclipxl_utils/text_encoder/__pycache__/__init__.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/compliance/videoclipxl_utils/text_encoder/__pycache__/model_text_encoder.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/compliance/videoclipxl_utils/text_encoder/__pycache__/model_text_encoder.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..e4a6eca0c471ee8df2bf017db20323a39d23c670 Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/compliance/videoclipxl_utils/text_encoder/__pycache__/model_text_encoder.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/compliance/videoclipxl_utils/text_encoder/__pycache__/simple_tokenizer.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/compliance/videoclipxl_utils/text_encoder/__pycache__/simple_tokenizer.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..e8cfc25afffced1b32a0a11dc9f0c3edea2718af Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/compliance/videoclipxl_utils/text_encoder/__pycache__/simple_tokenizer.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/compliance/videoclipxl_utils/text_encoder/__pycache__/text_encoder.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/compliance/videoclipxl_utils/text_encoder/__pycache__/text_encoder.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..af85cdded102c50a0ac1733fed43516335019e59 Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/compliance/videoclipxl_utils/text_encoder/__pycache__/text_encoder.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/compliance/videoclipxl_utils/vision_encoder/__pycache__/__init__.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/compliance/videoclipxl_utils/vision_encoder/__pycache__/__init__.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..c22de4792af511cc676cb05e27d0600fb5146122 Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/compliance/videoclipxl_utils/vision_encoder/__pycache__/__init__.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/compliance/videoclipxl_utils/vision_encoder/__pycache__/clip_vision.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/compliance/videoclipxl_utils/vision_encoder/__pycache__/clip_vision.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..a61b8c6d3866365a528a8bba660704805093d06a Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/compliance/videoclipxl_utils/vision_encoder/__pycache__/clip_vision.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/compliance/videoclipxl_utils/vision_encoder/__pycache__/model_vision_encoder.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/compliance/videoclipxl_utils/vision_encoder/__pycache__/model_vision_encoder.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..2d8fc95601566143cb2ad57517d7c5e21f8c77eb Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/compliance/videoclipxl_utils/vision_encoder/__pycache__/model_vision_encoder.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/fidelity/__pycache__/__init__.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/fidelity/__pycache__/__init__.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..c476b5686d0813d61ab2d8171231ef0383940b1c Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/fidelity/__pycache__/__init__.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/fidelity/__pycache__/content_fidelity.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/fidelity/__pycache__/content_fidelity.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..9da79dbca843f3888f5fc7f01b0953462c51b43c Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/fidelity/__pycache__/content_fidelity.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/fidelity/__pycache__/motion_fidelity.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/fidelity/__pycache__/motion_fidelity.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..d0c7fcf37c87497a050089987850b3fd9ecb997a Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/fidelity/__pycache__/motion_fidelity.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/fidelity/__pycache__/semantic_fidelity.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/fidelity/__pycache__/semantic_fidelity.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..10acb1edf249a84b05061f2d76446a23a2991b30 Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/fidelity/__pycache__/semantic_fidelity.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/fidelity/cotracker/__pycache__/__init__.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/fidelity/cotracker/__pycache__/__init__.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..85bac4dc1a70aaf81fbfb807c330039a183aebc4 Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/fidelity/cotracker/__pycache__/__init__.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/fidelity/cotracker/__pycache__/predictor.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/fidelity/cotracker/__pycache__/predictor.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..4cd6b14c8219973747a1bb7ec8804539b1f69578 Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/fidelity/cotracker/__pycache__/predictor.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/fidelity/cotracker/models/__pycache__/__init__.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/fidelity/cotracker/models/__pycache__/__init__.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..482450afb303af0479b4179a12cdb77267f04091 Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/fidelity/cotracker/models/__pycache__/__init__.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/fidelity/cotracker/models/__pycache__/build_cotracker.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/fidelity/cotracker/models/__pycache__/build_cotracker.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..6038b506d2c436cfd3c113eb478d52816124aef4 Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/fidelity/cotracker/models/__pycache__/build_cotracker.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/fidelity/cotracker/models/core/__pycache__/__init__.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/fidelity/cotracker/models/core/__pycache__/__init__.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..7664765eba5ff6138ca14c732db5591e069fc90a Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/fidelity/cotracker/models/core/__pycache__/__init__.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/fidelity/cotracker/models/core/__pycache__/embeddings.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/fidelity/cotracker/models/core/__pycache__/embeddings.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..89103712ab27af4c61eb044ae5651244ee5a5184 Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/fidelity/cotracker/models/core/__pycache__/embeddings.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/fidelity/cotracker/models/core/__pycache__/model_utils.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/fidelity/cotracker/models/core/__pycache__/model_utils.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..cc7a3c6bc88ca259909fbe3840a81f9225d32ca6 Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/fidelity/cotracker/models/core/__pycache__/model_utils.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/fidelity/cotracker/models/core/cotracker/__pycache__/__init__.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/fidelity/cotracker/models/core/cotracker/__pycache__/__init__.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..33db13721fe0770fe1353202f6d06c95138c4da2 Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/fidelity/cotracker/models/core/cotracker/__pycache__/__init__.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/fidelity/cotracker/models/core/cotracker/__pycache__/blocks.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/fidelity/cotracker/models/core/cotracker/__pycache__/blocks.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..1ac464abd9aaeba51788cfbd2a8fe76519e3efaa Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/fidelity/cotracker/models/core/cotracker/__pycache__/blocks.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/fidelity/cotracker/models/core/cotracker/__pycache__/cotracker.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/fidelity/cotracker/models/core/cotracker/__pycache__/cotracker.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..f8450002fc405e3f9ad7086be1883306ecd8cb35 Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/fidelity/cotracker/models/core/cotracker/__pycache__/cotracker.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/fidelity/cotracker/models/core/cotracker/__pycache__/cotracker3_offline.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/fidelity/cotracker/models/core/cotracker/__pycache__/cotracker3_offline.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..5d56f849b5b82c60fc975fd4020687e0aff21a95 Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/fidelity/cotracker/models/core/cotracker/__pycache__/cotracker3_offline.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/fidelity/cotracker/models/core/cotracker/__pycache__/cotracker3_online.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/fidelity/cotracker/models/core/cotracker/__pycache__/cotracker3_online.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..4b9b7f1a11bf07bd435a29e9bae467d79596ed9c Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/fidelity/cotracker/models/core/cotracker/__pycache__/cotracker3_online.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/fidelity/cotracker/utils/__pycache__/__init__.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/fidelity/cotracker/utils/__pycache__/__init__.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..bf9a12b1d2e1d7aaec778b5758c43e127754cf03 Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/fidelity/cotracker/utils/__pycache__/__init__.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/fidelity/cotracker/utils/__pycache__/visualizer.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/fidelity/cotracker/utils/__pycache__/visualizer.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..7f31f9d7df122a2094099b7f139e3b84f1fa3d69 Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/fidelity/cotracker/utils/__pycache__/visualizer.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/fidelity/qwen_vl_utils/__pycache__/__init__.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/fidelity/qwen_vl_utils/__pycache__/__init__.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..e736959262bc3ec2fb63a26db7724f0293f230e1 Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/fidelity/qwen_vl_utils/__pycache__/__init__.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/fidelity/qwen_vl_utils/__pycache__/vision_process.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/fidelity/qwen_vl_utils/__pycache__/vision_process.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..0965ee9bcf1b624bd72ffafe4589da560eef75d1 Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/fidelity/qwen_vl_utils/__pycache__/vision_process.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/fidelity/videoclipxl_utils/__pycache__/__init__.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/fidelity/videoclipxl_utils/__pycache__/__init__.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..b8c8e430d1d92c7fd8dd158b8a9da2f6f5675d9b Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/fidelity/videoclipxl_utils/__pycache__/__init__.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/fidelity/videoclipxl_utils/__pycache__/modeling.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/fidelity/videoclipxl_utils/__pycache__/modeling.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..b8f4b5c973ab7609d76a21661c4db278050be9b6 Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/fidelity/videoclipxl_utils/__pycache__/modeling.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/fidelity/videoclipxl_utils/text_encoder/__pycache__/__init__.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/fidelity/videoclipxl_utils/text_encoder/__pycache__/__init__.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..b5f0b23e97db33963700366d54fa78e51d924777 Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/fidelity/videoclipxl_utils/text_encoder/__pycache__/__init__.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/fidelity/videoclipxl_utils/text_encoder/__pycache__/model_text_encoder.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/fidelity/videoclipxl_utils/text_encoder/__pycache__/model_text_encoder.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..15a9c2a9baecf7f81c520b8cfb1e5812769f8508 Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/fidelity/videoclipxl_utils/text_encoder/__pycache__/model_text_encoder.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/fidelity/videoclipxl_utils/text_encoder/__pycache__/simple_tokenizer.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/fidelity/videoclipxl_utils/text_encoder/__pycache__/simple_tokenizer.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..193b6e3abbc11c926c2c2aa029521dce2a07c211 Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/fidelity/videoclipxl_utils/text_encoder/__pycache__/simple_tokenizer.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/fidelity/videoclipxl_utils/text_encoder/__pycache__/text_encoder.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/fidelity/videoclipxl_utils/text_encoder/__pycache__/text_encoder.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..b9f9b1ccf0b26136bbce9bc3401d07eab917c253 Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/fidelity/videoclipxl_utils/text_encoder/__pycache__/text_encoder.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/fidelity/videoclipxl_utils/vision_encoder/__pycache__/__init__.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/fidelity/videoclipxl_utils/vision_encoder/__pycache__/__init__.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..d2d2847394908c71005b7a538716e1994fd6c3c4 Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/fidelity/videoclipxl_utils/vision_encoder/__pycache__/__init__.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/fidelity/videoclipxl_utils/vision_encoder/__pycache__/clip_vision.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/fidelity/videoclipxl_utils/vision_encoder/__pycache__/clip_vision.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..7162991df659076f680f11c27d833a375010c5c3 Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/fidelity/videoclipxl_utils/vision_encoder/__pycache__/clip_vision.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/fidelity/videoclipxl_utils/vision_encoder/__pycache__/model_vision_encoder.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/fidelity/videoclipxl_utils/vision_encoder/__pycache__/model_vision_encoder.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..32125f7072df997c11f855b425521c33228e083c Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/fidelity/videoclipxl_utils/vision_encoder/__pycache__/model_vision_encoder.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/quality/__pycache__/__init__.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/quality/__pycache__/__init__.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..4fe94bf2d09c9df9a256a00a37f934194f9ef565 Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/quality/__pycache__/__init__.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/quality/__pycache__/background_consistency.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/quality/__pycache__/background_consistency.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..43dfdb81cfb072e092bc78321f985b12a16a11ae Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/quality/__pycache__/background_consistency.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/quality/__pycache__/motion_smoothness.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/quality/__pycache__/motion_smoothness.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..aff3753aaf68c383164ac77bb85b3b8f4d952d2b Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/quality/__pycache__/motion_smoothness.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/quality/__pycache__/subject_consistency.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/quality/__pycache__/subject_consistency.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..7f8fe6a0e2c95924fafdfd8ab43a5badb3dc5db2 Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/quality/__pycache__/subject_consistency.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/quality/__pycache__/temporal_flickering.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/quality/__pycache__/temporal_flickering.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..d6d87e9e504b4e08c91bf8402894f56f05a324b3 Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/quality/__pycache__/temporal_flickering.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/quality/__pycache__/vtss.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/quality/__pycache__/vtss.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..288bfbcaf6a7790978e29f74e9719c571db5f781 Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/quality/__pycache__/vtss.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/quality/amt/__pycache__/__init__.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/quality/amt/__pycache__/__init__.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..6720976d450b99c950775f154200311f35319907 Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/quality/amt/__pycache__/__init__.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/quality/amt/networks/__pycache__/AMT-G.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/quality/amt/networks/__pycache__/AMT-G.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..93ab12cf7c213e5c08ea84a29da8a8055f0c595c Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/quality/amt/networks/__pycache__/AMT-G.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/quality/amt/networks/__pycache__/__init__.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/quality/amt/networks/__pycache__/__init__.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..baa8c15c3ffa73c612d36f39d6ce77f150d191ec Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/quality/amt/networks/__pycache__/__init__.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/quality/amt/networks/blocks/__pycache__/__init__.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/quality/amt/networks/blocks/__pycache__/__init__.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..0848113872b4f9990ae77cb404e76f690455709b Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/quality/amt/networks/blocks/__pycache__/__init__.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/quality/amt/networks/blocks/__pycache__/feat_enc.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/quality/amt/networks/blocks/__pycache__/feat_enc.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..71bc7e21284fd901f9541aa3647b34eef2691ef5 Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/quality/amt/networks/blocks/__pycache__/feat_enc.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/quality/amt/networks/blocks/__pycache__/ifrnet.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/quality/amt/networks/blocks/__pycache__/ifrnet.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..a40b18a944ffc86b88d72c12fcbbb8078bafbe84 Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/quality/amt/networks/blocks/__pycache__/ifrnet.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/quality/amt/networks/blocks/__pycache__/multi_flow.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/quality/amt/networks/blocks/__pycache__/multi_flow.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..527b579723c839ddb55e41ee06d5285a0eca3ca8 Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/quality/amt/networks/blocks/__pycache__/multi_flow.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/quality/amt/networks/blocks/__pycache__/raft.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/quality/amt/networks/blocks/__pycache__/raft.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..d503c27bb07cde753ce838ebe9733e21f3aabba3 Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/quality/amt/networks/blocks/__pycache__/raft.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/quality/amt/utils/__pycache__/__init__.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/quality/amt/utils/__pycache__/__init__.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..3327d36244a69db0703c3fd9993fa8fbe0af819f Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/quality/amt/utils/__pycache__/__init__.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/quality/amt/utils/__pycache__/build_utils.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/quality/amt/utils/__pycache__/build_utils.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..cda8a8baf475cd35b8c8a49e70d940756ba8c388 Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/quality/amt/utils/__pycache__/build_utils.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/quality/amt/utils/__pycache__/flow_utils.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/quality/amt/utils/__pycache__/flow_utils.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..c6561adb890040dc8df1faeda34f2aa3a05425de Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/quality/amt/utils/__pycache__/flow_utils.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/quality/amt/utils/__pycache__/utils.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/quality/amt/utils/__pycache__/utils.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..2ccce72f1f6ec947ed22df6300445897f72ae108 Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/quality/amt/utils/__pycache__/utils.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/quality/training_suitability_assessment/__pycache__/__init__.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/quality/training_suitability_assessment/__pycache__/__init__.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..ee977a324dc738b6d383ef1adf966c74d5dc10d6 Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/quality/training_suitability_assessment/__pycache__/__init__.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/quality/training_suitability_assessment/__pycache__/conv_backbone.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/quality/training_suitability_assessment/__pycache__/conv_backbone.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..536d38e4ef00e103d7b92a6a9933bcd8e7312a59 Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/quality/training_suitability_assessment/__pycache__/conv_backbone.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/quality/training_suitability_assessment/__pycache__/datasets.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/quality/training_suitability_assessment/__pycache__/datasets.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..46028ef40b544e92256183b1251c87d706068093 Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/quality/training_suitability_assessment/__pycache__/datasets.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/quality/training_suitability_assessment/__pycache__/head.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/quality/training_suitability_assessment/__pycache__/head.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..0e7a9b6c842854643f31723c082614c21cc6b5ba Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/quality/training_suitability_assessment/__pycache__/head.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/quality/training_suitability_assessment/__pycache__/model.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/quality/training_suitability_assessment/__pycache__/model.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..e7d6e2a46f34e55c580a3254bafedbf8464d77e0 Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/quality/training_suitability_assessment/__pycache__/model.cpython-312.pyc differ diff --git a/benchmarks/edit/code/IVEBench/metrics/quality/training_suitability_assessment/__pycache__/swin_backbone.cpython-312.pyc b/benchmarks/edit/code/IVEBench/metrics/quality/training_suitability_assessment/__pycache__/swin_backbone.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..be32a5f291398feaf31c79984424c69bfedc9859 Binary files /dev/null and b/benchmarks/edit/code/IVEBench/metrics/quality/training_suitability_assessment/__pycache__/swin_backbone.cpython-312.pyc differ diff --git a/benchmarks/edit/run_traditional_metrics.py b/benchmarks/edit/run_traditional_metrics.py new file mode 100644 index 0000000000000000000000000000000000000000..745a27124b16236f2ba502381b08f6b174688dc0 --- /dev/null +++ b/benchmarks/edit/run_traditional_metrics.py @@ -0,0 +1,233 @@ +#!/usr/bin/env python3 +"""Run no-model and optional CLIP traditional metrics on six-method manifests.""" + +from __future__ import annotations + +import argparse +import csv +import json +import math +from pathlib import Path +from typing import Any + +import numpy as np +from PIL import Image + + +CSV_FIELDS = [ + "split", + "sample_id", + "method", + "instruction", + "source_video", + "edited_video", + "num_frames_sampled", + "pixel_mse", + "pixel_psnr", + "source_edit_l1", + "global_ssim", + "edited_frame_diff_mae", + "temporal_flicker_score", + "clip_t", + "clip_frame_consistency", + "clip_source_edit_similarity", + "error", +] + + +def read_jsonl(path: Path) -> list[dict[str, Any]]: + rows: list[dict[str, Any]] = [] + with path.open("r", encoding="utf-8") as handle: + for line in handle: + text = line.strip() + if text: + rows.append(json.loads(text)) + return rows + + +def uniform_indices(length: int, count: int) -> list[int]: + if length <= 0: + return [] + if count <= 0 or count >= length: + return list(range(length)) + if count == 1: + return [length // 2] + values: list[int] = [] + for index in range(count): + value = round(index * (length - 1) / (count - 1)) + if not values or value != values[-1]: + values.append(value) + return values + + +def load_video_frames(path: Path, count: int, size: int | None) -> list[Image.Image]: + if not path.exists(): + raise FileNotFoundError(path) + try: + import imageio.v3 as iio + except Exception as exc: + raise RuntimeError("Please install imageio and imageio-ffmpeg to decode mp4 files") from exc + + frames = [Image.fromarray(frame).convert("RGB") for frame in iio.imiter(str(path))] + if not frames: + raise RuntimeError(f"no frames decoded from {path}") + sampled = [frames[i] for i in uniform_indices(len(frames), count)] + if size: + sampled = [frame.resize((size, size), Image.Resampling.BICUBIC) for frame in sampled] + return sampled + + +def frames_to_float(frames: list[Image.Image]) -> np.ndarray: + return np.stack([np.asarray(frame, dtype=np.float32) for frame in frames], axis=0) + + +def align_frame_arrays(source: np.ndarray, edited: np.ndarray) -> tuple[np.ndarray, np.ndarray]: + n = min(len(source), len(edited)) + if n <= 0: + raise RuntimeError("no aligned frames") + return source[:n], edited[:n] + + +def global_ssim_gray(a: np.ndarray, b: np.ndarray) -> float: + """Simple global SSIM over all sampled grayscale pixels, no skimage dependency.""" + a_gray = 0.299 * a[..., 0] + 0.587 * a[..., 1] + 0.114 * a[..., 2] + b_gray = 0.299 * b[..., 0] + 0.587 * b[..., 1] + 0.114 * b[..., 2] + x = a_gray.reshape(-1).astype(np.float64) + y = b_gray.reshape(-1).astype(np.float64) + mux = x.mean() + muy = y.mean() + vx = ((x - mux) ** 2).mean() + vy = ((y - muy) ** 2).mean() + cov = ((x - mux) * (y - muy)).mean() + c1 = (0.01 * 255) ** 2 + c2 = (0.03 * 255) ** 2 + return float(((2 * mux * muy + c1) * (2 * cov + c2)) / ((mux**2 + muy**2 + c1) * (vx + vy + c2))) + + +def no_model_metrics(source_frames: list[Image.Image], edited_frames: list[Image.Image]) -> dict[str, float | int]: + source = frames_to_float(source_frames) + edited = frames_to_float(edited_frames) + source, edited = align_frame_arrays(source, edited) + + diff = edited - source + mse = float(np.mean(diff**2)) + psnr = float("inf") if mse == 0 else float(20 * math.log10(255.0 / math.sqrt(mse))) + l1 = float(np.mean(np.abs(diff)) / 255.0) + ssim = global_ssim_gray(source, edited) + + if len(edited) >= 2: + frame_diff = np.abs(edited[1:] - edited[:-1]) + frame_diff_mae = float(np.mean(frame_diff)) + flicker_score = float(max(0.0, min(1.0, (255.0 - frame_diff_mae) / 255.0))) + else: + frame_diff_mae = 0.0 + flicker_score = 1.0 + + return { + "num_frames_sampled": int(len(edited)), + "pixel_mse": mse, + "pixel_psnr": psnr, + "source_edit_l1": l1, + "global_ssim": ssim, + "edited_frame_diff_mae": frame_diff_mae, + "temporal_flicker_score": flicker_score, + } + + +class ClipMetrics: + def __init__(self, model_dir: Path, device: str, batch_size: int) -> None: + import torch + from transformers import CLIPModel, CLIPProcessor + + self.torch = torch + self.device = device + self.batch_size = batch_size + self.processor = CLIPProcessor.from_pretrained(str(model_dir)) + self.model = CLIPModel.from_pretrained(str(model_dir)).to(device) + self.model.eval() + + def image_features(self, frames: list[Image.Image]): + chunks = [] + with self.torch.inference_mode(): + for start in range(0, len(frames), self.batch_size): + batch = frames[start : start + self.batch_size] + inputs = self.processor(images=batch, return_tensors="pt").to(self.device) + feats = self.model.get_image_features(**inputs) + feats = feats / feats.norm(dim=-1, keepdim=True).clamp_min(1e-12) + chunks.append(feats) + return self.torch.cat(chunks, dim=0) + + def text_feature(self, text: str): + with self.torch.inference_mode(): + inputs = self.processor(text=[text], return_tensors="pt", padding=True, truncation=True).to(self.device) + feat = self.model.get_text_features(**inputs) + feat = feat / feat.norm(dim=-1, keepdim=True).clamp_min(1e-12) + return feat[0] + + def compute(self, source_frames: list[Image.Image], edited_frames: list[Image.Image], instruction: str) -> dict[str, float | None]: + source_n = min(len(source_frames), len(edited_frames)) + source_frames = source_frames[:source_n] + edited_frames = edited_frames[:source_n] + source_feat = self.image_features(source_frames) + edited_feat = self.image_features(edited_frames) + text_feat = self.text_feature(instruction) + clip_t = float((edited_feat @ text_feat).mean().item() * 100.0) + clip_frame = None + if edited_feat.shape[0] >= 2: + clip_frame = float((edited_feat[:-1] * edited_feat[1:]).sum(dim=-1).mean().item() * 100.0) + source_edit = float((source_feat * edited_feat).sum(dim=-1).mean().item() * 100.0) + return { + "clip_t": clip_t, + "clip_frame_consistency": clip_frame, + "clip_source_edit_similarity": source_edit, + } + + +def empty_result(row: dict[str, Any]) -> dict[str, Any]: + result = {field: "" for field in CSV_FIELDS} + for key in ("split", "sample_id", "method", "instruction", "source_video", "edited_video"): + result[key] = row.get(key, "") + return result + + +def main() -> None: + parser = argparse.ArgumentParser() + parser.add_argument("--manifest", type=Path, required=True) + parser.add_argument("--output", type=Path, required=True) + parser.add_argument("--frames-per-video", type=int, default=16) + parser.add_argument("--resize", type=int, default=256, help="Resize sampled frames to square size before metrics; 0 disables.") + parser.add_argument("--clip-model-dir", type=Path, default=None) + parser.add_argument("--device", default="cpu") + parser.add_argument("--clip-batch-size", type=int, default=8) + args = parser.parse_args() + + rows = read_jsonl(args.manifest) + args.output.parent.mkdir(parents=True, exist_ok=True) + + clip = None + if args.clip_model_dir: + clip = ClipMetrics(args.clip_model_dir, args.device, args.clip_batch_size) + + with args.output.open("w", encoding="utf-8", newline="") as handle: + writer = csv.DictWriter(handle, fieldnames=CSV_FIELDS) + writer.writeheader() + for index, row in enumerate(rows, start=1): + result = empty_result(row) + try: + source_video = Path(str(row["source_video"])) + edited_video = Path(str(row["edited_video"])) + source_frames = load_video_frames(source_video, args.frames_per_video, args.resize or None) + edited_frames = load_video_frames(edited_video, args.frames_per_video, args.resize or None) + result.update(no_model_metrics(source_frames, edited_frames)) + if clip is not None: + result.update(clip.compute(source_frames, edited_frames, str(row["instruction"]))) + except Exception as exc: + result["error"] = repr(exc) + writer.writerow(result) + if index % 20 == 0: + print(f"processed {index}/{len(rows)}") + print(f"wrote {len(rows)} rows -> {args.output}") + + +if __name__ == "__main__": + main() diff --git a/benchmarks/edit/traditional_eval_notes.md b/benchmarks/edit/traditional_eval_notes.md new file mode 100644 index 0000000000000000000000000000000000000000..f17aee45fbdf1cdeca693d27e35d6e66bbd54456 --- /dev/null +++ b/benchmarks/edit/traditional_eval_notes.md @@ -0,0 +1,226 @@ +# Traditional / Pretrained Metric Plan for Our 6 Edit Results + +目标数据： + +- `val20`: `out/edit_model_face_stage1/eval_outputs/{wan_only,ditto_global,full,text,vace_hint,vace_context}/val_XXXX.mp4` +- `val100`: `out/edit_model_face_stage1/eval_outputs_val100/{wan_only,ditto_global,full,text,vace_hint,vace_context}/val_XXXX.mp4` +- sample metadata: + - `out/edit_model_face_stage1/eval_samples/val_20.jsonl` + - `out/edit_model_face_stage1/eval_samples/val_100.jsonl` + +## No-Model Metrics: Can Run Immediately + +这些不需要下载模型，只需要 Python 包能读视频。脚本会用 `imageio`/`PIL`/`numpy`。 + +| Metric | 对应 benchmark 口径 | 输入 | 分数方向 | +|---|---|---|---| +| `pixel_mse` | FiVE MSE / 结构差异近似 | source video + edited video | 低更好 | +| `pixel_psnr` | FiVE PSNR | source video + edited video | 高更好 | +| `global_ssim` | FiVE/VEditBench SSIM 近似 | source video + edited video | 高更好 | +| `temporal_flicker_score` | IVE Temporal Flickering 近似 | edited video | 高更好；这里按 `(255 - frame_diff_mae) / 255` | +| `edited_frame_diff_mae` | 时间闪烁原始量 | edited video | 低更好 | +| `source_edit_l1` | Structure distance / non-mask preservation 近似 | source video + edited video | 低更好 | +| `num_frames_sampled` | 记录采样帧数 | edited video | 非质量分 | + +注意： + +- 这些是全图指标，不等同 FiVE 官方 masked background preservation。没有 mask 时，不能说是正式 FiVE background preservation。 +- SSIM 这里实现的是全局灰度 SSIM 近似，不依赖 `skimage`；如果要官方/库级 SSIM，可后续换成 `skimage.metrics.structural_similarity`。 + +## Model-Dependent Traditional / Pretrained Metrics + +下载目录统一使用： + +```bash +cd /Users/ouzhang/Desktop/low-high/low-high-new +mkdir -p models +``` + +### CLIP + +用途： + +- FiVE `CLIPSIM` / `CLIPS.edit` +- EditBoard `success_rate`, `clip_similarity`, `background_consistency` +- 我们脚本里的 `clip_t`, `clip_frame_consistency`, `clip_source_edit_similarity` + +下载： + +```bash +hf download openai/clip-vit-large-patch14 \ + --local-dir models/openai_clip-vit-large-patch14 +``` + +可选：EditBoard 原代码还会用 OpenAI CLIP `ViT-B/32` / `ViT-L/14`，如果跑官方 EditBoard，需要让 `clip` 包能下载或提前缓存对应权重。 + +### DINO + +用途： + +- EditBoard `subject_consistency` +- IVE-Bench `subject_consistency` + +下载/缓存方式通常是 torch hub： + +```bash +python3 - <<'PY' +import torch +torch.hub.load('facebookresearch/dino:main', 'dino_vitb16') +torch.hub.load('facebookresearch/dino:main', 'dino_vits16') +PY +``` + +如果真实机无网，需在有网机器缓存 `~/.cache/torch/hub` 后复制过去。 + +### LAION Aesthetic Predictor + +用途： + +- EditBoard `aesthetic_quality` + +下载： + +```bash +mkdir -p models/laion_aesthetic +wget -O models/laion_aesthetic/sa_0_4_vit_l_14_linear.pth \ + https://github.com/LAION-AI/aesthetic-predictor/raw/main/sa_0_4_vit_l_14_linear.pth +``` + +### MUSIQ / pyiqa + +用途： + +- EditBoard `imaging_quality` + +安装/缓存： + +```bash +python3 -m pip install pyiqa +python3 - <<'PY' +import pyiqa +pyiqa.create_metric('musiq', device='cpu') +PY +``` + +### LPIPS + +用途： + +- FiVE background preservation `LPIPS` + +安装： + +```bash +python3 -m pip install lpips +``` + +或使用 `torchmetrics.image.lpip.LearnedPerceptualImagePatchSimilarity`，它会按需拉取 backbone 权重。 + +### CoTracker3 + +用途： + +- FiVE `Motion Fidelity Score` +- IVE `Motion Fidelity` +- VEditBench `Motion Similarity` + +下载： + +```bash +mkdir -p models/cotracker3 +hf download facebook/cotracker3 scaled_offline.pth \ + --local-dir models/cotracker3 +hf download facebook/cotracker3 baseline_offline.pth \ + --local-dir models/cotracker3 +``` + +### VideoCLIP-XL-v2 + +用途： + +- IVE `OSC`, `PSC`, `SF` + +下载： + +```bash +hf download alibaba-pai/VideoCLIP-XL-v2 \ + --local-dir models/VideoCLIP-XL-v2 +``` + +### GroundingDINO + +用途： + +- IVE `Quantity Accuracy` + +下载： + +```bash +mkdir -p models/GroundingDINO +wget -O models/GroundingDINO/groundingdino_swinb_cogcoor.pth \ + https://github.com/IDEA-Research/GroundingDINO/releases/download/v0.1.0-alpha2/groundingdino_swinb_cogcoor.pth +``` + +还需要 GroundingDINO 代码和 config： + +```bash +git clone https://github.com/IDEA-Research/GroundingDINO.git models/GroundingDINO/code +``` + +### AMT-G + +用途： + +- IVE / VEditBench `Motion Smoothness` + +下载位置按 IVE `metrics/path.yml` 配置到 `amt-g.pth`。如果使用作者给出的 HF/项目链接，放到： + +```bash +mkdir -p models/amt +# 根据 AMT 官方或 IVE README 给出的链接下载 amt-g.pth 到： +# models/amt/amt-g.pth +``` + +### Q-Align + +用途： + +- VEditBench `Image Quality`, `Image Aesthetic`, `Video Quality` + +本地 reference 没有完整 VEditBench official eval code。需要另配 Q-Align 官方 scorer 和权重，下载后再接入统一 manifest。 + +## Scripts Added Here + +- `build_six_method_manifest.py`: 为 val20/val100 构建统一 manifest。 +- `run_traditional_metrics.py`: 默认跑 no-model metrics；如果传入 `--clip-model-dir models/openai_clip-vit-large-patch14`，额外跑 CLIP 指标。 + +示例： + +```bash +cd /Users/ouzhang/Desktop/low-high/low-high-new + +python3 reference/benchmarks/edit/build_six_method_manifest.py \ + --repo-root . \ + --output-dir out/edit_model_face_stage1/traditional_eval_manifests + +python3 reference/benchmarks/edit/run_traditional_metrics.py \ + --manifest out/edit_model_face_stage1/traditional_eval_manifests/val20.jsonl \ + --output out/edit_model_face_stage1/traditional_eval_metrics/val20_metrics.csv \ + --frames-per-video 16 + +python3 reference/benchmarks/edit/run_traditional_metrics.py \ + --manifest out/edit_model_face_stage1/traditional_eval_manifests/val100.jsonl \ + --output out/edit_model_face_stage1/traditional_eval_metrics/val100_metrics.csv \ + --frames-per-video 16 +``` + +带 CLIP： + +```bash +python3 reference/benchmarks/edit/run_traditional_metrics.py \ + --manifest out/edit_model_face_stage1/traditional_eval_manifests/val100.jsonl \ + --output out/edit_model_face_stage1/traditional_eval_metrics/val100_metrics_clip.csv \ + --frames-per-video 16 \ + --clip-model-dir models/openai_clip-vit-large-patch14 \ + --device cuda +``` diff --git "a/benchmarks/edit/\350\257\204\346\265\213\350\241\250\346\240\274.md" "b/benchmarks/edit/\350\257\204\346\265\213\350\241\250\346\240\274.md" new file mode 100644 index 0000000000000000000000000000000000000000..1e6c7f87719e84a90504617bbab258f55bb758a7 --- /dev/null +++ "b/benchmarks/edit/\350\257\204\346\265\213\350\241\250\346\240\274.md" @@ -0,0 +1,133 @@ +# 视频编辑评测 Benchmark 方法核对表 + +本表基于 `reference/benchmarks/edit` 下的论文提取文本、README 和代码整理。最后一列给出“如何落地到我们的 6 组结果”的最小实现方式。 +我们的 6 组结果建议统一成 manifest：`sample_id, method, source_video, edited_video, instruction, source_prompt, target_prompt, mask_path, target_phrase, target_span, category`。其中 `mask_path/target_phrase/target_span/category` 不是所有 benchmark 都需要。 + +重要区分： + +- 如果只评测现有 `val20/val100` 的 `wan_only, ditto_global, full, text, vace_hint, vace_context`，得到的是“内部横向比较”。 +- 如果要和 benchmark 论文或 leaderboard 可比，必须用对应 benchmark 的原始视频、prompt、mask/metadata 重新生成 6 组结果。 + +评测类型口径： + +- `VLM-as-judge`：直接把视频和指令喂给 VLM，让它按 prompt/rubric 判断或打分。重点检查代码里是否有 system/user prompt。 +- `传统/预训练指标`：不用在该 benchmark 数据上重新训练评估器，主要调用 CLIP、DINO、CoTracker、GroundingDINO、VideoCLIP、Q-Align、AMT、LPIPS/SSIM/PSNR/NIQE 等现成模型或客观公式。重点列出需要下载的模型。 +- `训练型 reward/evaluator`：作者在自己数据集和人工分上训练出的评估器/reward model，例如 VE-Bench QA、VEFX-Reward、IVE 的 VTSS。这类不是你当前重点，但表里保留。 + +## 总览 + +| Benchmark | 官方代码/资料 | 是否适合直接评测现有 6 组 val 输出 | 主要额外依赖 | +|---|---|---:|---| +| VE-Bench | `code/VE-Bench`，`pdf/_extracted/ve-bench.txt` | 是，但依赖作者 QA checkpoint | `vebench` / Google Drive checkpoints | +| EditBoard | `code/EditBoard` | 部分是；`semantic_score` 需要 mask | CLIP、DINO、LAION aesthetic、MUSIQ/pyiqa | +| FiVE | `code/FiVE-Bench`，`pdf/_extracted/five.txt` | 部分是；正式 FiVE 需要 FiVE 数据和 masks | FiVE dataset、CLIP、CoTracker3、Qwen2.5-VL-7B、pyiqa | +| OpenVE-3M / OpenVE-Bench | `code/OpenVE-3M/OpenVE-Bench` | 可用 VLM judge 适配；正式可比需 OpenVE-Bench 数据 | Seed/Gemini API 或 InternVL/Qwen3VL 本地模型 | +| IVE-Bench | `code/IVEBench`，`pdf/_extracted/ive-bench.txt` | 不建议直接全量适配；metadata 缺失 | VideoCLIP-XL-v2、Qwen2.5-VL-72B、CoTracker3、GroundingDINO、AMT-G、VTSS | +| VEFX-Bench | `code/VEFX-Bench`，`pdf/_extracted/vefx-bench.txt` | 是，最适合先跑 | `xiangbog/VEFX-Reward-4B` | +| VEditBench | `pdf/_extracted/vedit-bench.txt`，本地无完整官方 eval code | 可复现部分指标，但工程量大 | CLIP、ViCLIP、CoTracker、Q-Align、AMT/VBench、VideoMAE-v2 | + +## 逐项核对表 + +| Benchmark | 一级维度 | 二级指标 | 评测类型 | 评测一次用多少个 sample | 输入形态 | 分数制/方向 | VLM prompt / 传统指标需下载模型 | 原表描述核对 | 需要下载/配置 | 实现方式、代码命令、数据下载命令 | +|---|---|---|---|---|---|---|---|---|---|---| +| VE-Bench | Subjective-aligned QA | 总分 / Overall QA Score | 训练型 reward/evaluator | 单条 pair；也可批量；DB 含 28,080 human scores。 | 直接输入 mp4：src + dst + prompt。 | 连续回归分；训练监督来自主观分，README 提醒输出不是绝对 1-10；高更好。 | 不属于前两类重点；需 VE-Bench QA checkpoints。 | 需要拆开看。官方 `VEBenchModel().evaluate(prompt, src, dst)` 对外通常返回一个综合 human-aligned 分数；论文中这个综合分来自文本-视频一致性、源-目标关联性、视觉质量等分支融合，不是零样本 CLIP/FVD/LPIPS。 | 作者 checkpoints。README 说可 `pip install vebench`，本地推理需要从 Google Drive 下载所有 checkpoints 到 `ckpts`。 | 数据下载：`# VE-Bench DB: 见 README 的 baidu/google drive 链接`。安装：`pip install vebench`。单条：`python -m infer.py --single_test --src_path SRC.mp4 --dst_path EDIT.mp4 --prompt "PROMPT"`。Python API：`from vebench import VEBenchModel; VEBenchModel().evaluate(prompt, src, dst)`。对我们的 6 组结果：批量遍历 manifest，每行调用一次。 | +| VE-Bench | Text-Video Alignment | 文本-视频一致性 | 训练型 reward/evaluator | 单条 pair；也可批量；DB 含 28,080 human scores。 | 直接输入 mp4：src + dst + prompt。 | 连续回归子分支；官方默认不一定单独输出；高更好。 | 不属于前两类重点；需 VE-Bench text branch/BLIP+Temporal Adapter checkpoint。 | 原表正确。论文中使用 BLIP 视觉编码器和文本编码器，并加入 Temporal Adapter 建模视频时序，再接回归头预测文本-视频一致性；这是在 VE-Bench DB 人工主观分上训练出来的评估分支，不是直接用 CLIP cosine。 | 需要 VE-Bench 的 text 分支 checkpoint；本地代码配置在 `code/VE-Bench/vebench/configs/text.yaml`，整体 checkpoints 放 `ckpts`。 | 官方公开 API 不一定单独输出该子分支；若只用 `VEBenchModel().evaluate(...)`，拿到的是融合总分。若要单独导出，需要改 `code/VE-Bench/vebench/evaluator.py` 或底层 `infer.py`，把 text branch logits/score 写出。输入仍是 `prompt, SRC.mp4, EDIT.mp4`。 | +| VE-Bench | Source-Target Relationship / Relevance | 源视频-目标视频关联性 | 训练型 reward/evaluator | 单条 pair；也可批量；DB 含 28,080 human scores。 | 直接输入 mp4：src + dst + prompt。 | 连续回归子分支；官方默认不一定单独输出；高更好。 | 不属于前两类重点；需 VE-Bench source-target/Uniformer branch checkpoint。 | 原表基本正确。论文中将源视频和编辑后视频送入时空编码器，最终选择 Uniformer 相关 backbone，再拼接特征经 FFN 预测源-目标相关性；这是训练型评估器。 | 需要 VE-Bench source-target/relevance 分支 checkpoint；整体 checkpoints 来自 README Google Drive。 | 官方单条命令同总分：`python -m infer.py --single_test --src_path SRC.mp4 --dst_path EDIT.mp4 --prompt "PROMPT"`。如果要单独记录该子项，需要在 evaluator 内部暴露 source-target branch 输出；对我们的 6 组结果仍按 manifest 每行调用。 | +| VE-Bench | Visual Quality | 编辑后视频质量 | 训练型 reward/evaluator | 单条 pair；也可批量；DB 含 28,080 human scores。 | 直接输入 mp4：src + dst + prompt。 | 连续回归子分支；官方默认不一定单独输出；高更好。 | 不属于前两类重点；需 VE-Bench DOVER/ConvNeXt/VideoSwin visual branch checkpoint。 | 原表正确。论文中视觉质量分支初始化自 DOVER：审美分支为 ConvNeXt，预训练于 AVA；失真/技术质量分支为 VideoSwin，预训练于 GRPB；之后在 VE-Bench DB 上继续训练回归头和分支参数。 | 需要 DOVER/visual quality 相关 checkpoint，随 VE-Bench checkpoints 放入 `ckpts`；依赖来自 `code/VE-Bench/requirements.txt`。 | 官方 API 默认融合输出总分；若需要单独 Visual Quality 分，需改 evaluator 导出 visual branch score。命令仍是 `python -m infer.py --single_test --src_path SRC.mp4 --dst_path EDIT.mp4 --prompt "PROMPT"`。若只想要替代视觉质量指标，可以先跑 EditBoard `aesthetic_quality/imaging_quality` 或 VEFX `RQ`。 | +| EditBoard | Fidelity | FF-alpha | 传统/预训练指标 | 单条 pair 或脚本批量；我们 val20=120、val100=600。 | 先拆帧：source/edit 帧目录，通常 512x512。 | 连续误差值，无固定上限；低更好。 | 无需下载深度模型；依赖光流/warping 代码和视频帧。 | 基本正确。README 和代码把它作为 `ff_alpha` 维度；需要原视频帧目录和编辑视频帧目录。它是基于光流/warping/像素误差的手工 fidelity 指标。 | 不需要大 VLM；需要 EditBoard requirements。光流相关依赖由仓库代码/requirements 负责。 | 安装：`cd reference/benchmarks/edit/code/EditBoard && pip install -r requirements.txt`。预处理：`python preprocess.py --input_path /path/to/videos --output_path /path/to/frames`。评测：`python -W ignore evaluate.py --dimension ff_alpha --original_video_path SRC_FRAMES --edited_video_path EDIT_FRAMES --output_path OUT --result_name NAME`。 | +| EditBoard | Fidelity | FF-beta | 传统/预训练指标 | 单条 pair 或脚本批量；我们 val20=120、val100=600。 | 先拆帧：source/edit 帧目录，通常 512x512。 | 平均 `1-cos(theta)`，理论约 0-2；低更好。 | 无需下载深度模型；依赖光流计算。 | 基本正确。README 中为 `ff_beta`；原表“比较光流方向夹角”与论文描述一致，属于手工运动/结构保真指标。 | 同 FF-alpha。 | 同 FF-alpha，把 `--dimension ff_beta` 或 `--dimension ff_alpha ff_beta`。 | +| EditBoard | Fidelity | Semantic Score | 传统/预训练指标 | 单条 pair 或脚本批量；我们 val20=120、val100=600。 | 先拆帧：source/edit/mask 帧目录，mask 必需。 | mask 区域像素差/误差，通常 0-1 附近；低更好。 | 无需下载评估模型；必须准备 semantic mask，可用 SAM2/Grounded-SAM 另行生成。 | 有条件正确。它确实使用 semantic mask 评估非编辑区域/语义区域的保持，但对我们的现有 val 输出没有现成 mask。没有 mask 不应硬跑。 | 必须有与 source/edit 帧数一致的 `semantic_mask_path`。可用 benchmark mask，或额外用 SAM2/Grounded-SAM 生成。 | 数据准备：为每个 sample 准备 `SRC_FRAMES, EDIT_FRAMES, MASK_FRAMES`，三者帧数一致。命令：`python -W ignore evaluate.py --dimension semantic_score --original_video_path SRC_FRAMES --edited_video_path EDIT_FRAMES --semantic_mask_path MASK_FRAMES --output_path OUT --result_name NAME`。 | +| EditBoard | Execution | Success Rate | 传统/预训练指标 | 单条 pair 或脚本批量；我们 val20=120、val100=600。 | 先拆帧：edited 帧目录 + source/target prompt。 | 0-1 比例；高更好。 | 需 CLIP ViT-B/32。 | 正确。代码维度为 `success_rate`，使用 CLIP 比较编辑帧与 source/target prompt 的相对匹配。 | CLIP，通常由 requirements 安装 `clip` 并自动/本地加载 ViT-B/32。 | 命令：`python -W ignore evaluate.py --dimension success_rate --edited_video_path EDIT_FRAMES --source_prompt "SOURCE" --target_prompt "TARGET" --output_path OUT --result_name NAME`。对我们的数据：如果只有 edit instruction，没有 source/target prompt，要先构造 `source_prompt/target_prompt`，否则 Success Rate 语义不完整。 | +| EditBoard | Execution | CLIP Similarity | 传统/预训练指标 | 单条 pair 或脚本批量；我们 val20=120、val100=600。 | 先拆帧：edited 帧目录 + source/target prompt。 | CLIP 相似度/概率连续值；高更好。 | 需 CLIP ViT-B/32。 | 正确。代码维度为 `clip_similarity`，计算编辑帧与 target prompt 的 CLIP cosine/相似度。 | CLIP。 | 命令：`python -W ignore evaluate.py --dimension clip_similarity --edited_video_path EDIT_FRAMES --source_prompt "SOURCE" --target_prompt "TARGET" --output_path OUT --result_name NAME`。 | +| EditBoard | Consistency | Subject Consistency | 传统/预训练指标 | 单条 pair 或脚本批量；我们 val20=120、val100=600。 | 先拆帧：edited 帧目录，通常 512x512。 | DINO 跨帧 cosine 连续值；高更好。 | 需 DINO ViT-B/16，代码通过 torch hub 加载。 | 正确。代码维度为 `subject_consistency`，用 DINO 特征做跨帧一致性。 | DINO，代码会通过 torch hub/本地缓存加载。真实机无网时要提前缓存模型。 | 命令：`python -W ignore evaluate.py --dimension subject_consistency --edited_video_path EDIT_FRAMES --output_path OUT --result_name NAME`。 | +| EditBoard | Consistency | Background Consistency | 传统/预训练指标 | 单条 pair 或脚本批量；我们 val20=120、val100=600。 | 先拆帧：edited 帧目录，通常 512x512。 | CLIP image 跨帧 cosine 连续值；高更好。 | 需 CLIP image encoder，通常 CLIP ViT-B/32。 | 正确。代码维度为 `background_consistency`，用 CLIP image feature 评估跨帧背景一致性。 | CLIP。 | 命令：`python -W ignore evaluate.py --dimension background_consistency --edited_video_path EDIT_FRAMES --output_path OUT --result_name NAME`。 | +| EditBoard | Style | Aesthetic Quality | 传统/预训练指标 | 单条 pair 或脚本批量；我们 val20=120、val100=600。 | 先拆帧：edited 帧目录，通常 512x512。 | LAION aesthetic 分数除以 10，约 0-1；高更好。 | 需 CLIP ViT-L/14 + LAION aesthetic predictor `sa_0_4_vit_l_14_linear.pth`。 | 正确。代码维度为 `aesthetic_quality`，用 LAION aesthetic predictor。 | LAION aesthetic predictor 权重，例如 `sa_0_4_vit_l_14_linear.pth`；CLIP ViT-L/14。 | 下载：`wget -O models/sa_0_4_vit_l_14_linear.pth https://github.com/LAION-AI/aesthetic-predictor/raw/main/sa_0_4_vit_l_14_linear.pth`。评测：`python -W ignore evaluate.py --dimension aesthetic_quality --edited_video_path EDIT_FRAMES --output_path OUT --result_name NAME`。 | +| EditBoard | Style | Imaging Quality | 传统/预训练指标 | 单条 pair 或脚本批量；我们 val20=120、val100=600。 | 先拆帧：edited 帧目录，通常 512x512。 | MUSIQ/technical quality 分数除以 100，约 0-1；高更好。 | 需 pyiqa/MUSIQ checkpoint，通常运行时下载或提前缓存。 | 基本正确。代码维度为 `imaging_quality`，README 归入九个维度；实现通常依赖 MUSIQ/pyiqa。 | pyiqa/MUSIQ 权重，可能运行时下载；无网环境需提前缓存。 | 安装：`pip install pyiqa`。评测：`python -W ignore evaluate.py --dimension imaging_quality --edited_video_path EDIT_FRAMES --output_path OUT --result_name NAME`。 | +| FiVE | Conventional Metrics | Structure Distance | 传统/预训练指标 | 官方 420 pairs；我们可批量 val20=120、val100=600。 | 多用抽帧/图片目录；视频需先抽帧。 | Structure Dist.×10^3；低更好。 | 通常无需额外模型；按帧/结构差异计算。 | 正确但原表过简。论文表格把 Structure Dist. 作为传统结构/差异指标，代码统一在 `evaluation/evaluate.py` 和 `metrics_calculator.py` 中计算。 | FiVE evaluation 环境。正式可比需要 FiVE 数据。 | 数据下载：`huggingface-cli download --repo-type dataset LIMinghan/FiVE-Fine-Grained-Video-Editing-Benchmark --local-dir FiVE_Bench/data`。评测入口：`cd reference/benchmarks/edit/code/FiVE-Bench && sh scripts/eval_FiVE.sh`，或直接改 `config.yaml` 后运行 `python evaluation/evaluate.py`。 | +| FiVE | Conventional Metrics | Background Preservation: PSNR / LPIPS / MSE / SSIM | 传统/预训练指标 | 官方 420 pairs；我们可批量 val20=120、val100=600。 | 多用抽帧/图片目录 + mask；视频需先抽帧。 | PSNR dB 高更好；LPIPS/MSE 低更好；SSIM 0-1 高更好。 | PSNR/MSE/SSIM 无需模型；LPIPS 需 torchmetrics/LPIPS backbone；还需要 bmasks。 | 正确。README 明确写在 editing mask 外计算 background preservation，含 PSNR、LPIPS、MSE、SSIM。 | FiVE `bmasks`；torchmetrics/LPIPS。 | 如果用 FiVE 正式数据：用其 `bmasks`。如果用我们的 val20/100：必须给每个 sample 生成 mask，否则 background preservation 只能退化成全图 PSNR/LPIPS/MSE/SSIM，不能等同 FiVE。 | +| FiVE | Text Alignment | CLIPSIM / CLIPS.edit | 传统/预训练指标 | 官方 420 pairs；我们可批量 val20=120、val100=600。 | 多用抽帧/图片目录 + mask；视频需先抽帧。 | CLIPScore 连续值；论文表通常约 20-30；高更好。 | 需 `openai/clip-vit-large-patch14`；`CLIPS.edit` 还需要编辑区域 mask。 | 正确。README 明确包括 full image 和 masked image 的 CLIP similarity。代码里使用 `openai/clip-vit-large-patch14`。 | `openai/clip-vit-large-patch14`；masked 版本需要 mask。 | 下载：`hf download openai/clip-vit-large-patch14 --local-dir models/openai_clip-vit-large-patch14`。正式评测：配置 FiVE `config.yaml` 后跑 `sh scripts/eval_FiVE.sh`。我们的现有数据：全图 CLIP 可直接做；`CLIPS.edit` 需要编辑区域 mask。 | +| FiVE | IQA | NIQE | 传统/预训练指标 | 官方 420 pairs；我们可批量 val20=120、val100=600。 | 多用抽帧/图片目录；视频需先抽帧。 | NIQE 无参考质量分；低更好。 | 需 pyiqa/IQA-PyTorch 的 NIQE 实现，通常无大模型权重。 | 正确。README 明确使用 NIQE，来自 IQA-PyTorch/pyiqa。 | `pyiqa`。 | 安装：`pip install pyiqa`。在 FiVE 官方流程里由 `evaluation/metrics_calculator.py` 调用；对我们的数据也可单独逐帧跑 NIQE 后平均。 | +| FiVE | Temporal Consistency | Motion Fidelity Score | 传统/预训练指标 | 官方 420 pairs；我们可批量 val20=120、val100=600。 | 视频或帧序列；CoTracker 内部取帧/轨迹。 | Motion Fidelity Score 连续值；论文表为 ×10^2；高更好。 | 需 CoTracker3 `scaled_offline.pth`。 | 正确。README 写 MFS 沿用 diffusion-motion-transfer 的 motion fidelity score；代码配置 CoTracker。 | `facebook/cotracker3/scaled_offline.pth`；安装 co-tracker。 | 下载：`wget -O checkpoints/scaled_offline.pth https://huggingface.co/facebook/cotracker3/resolve/main/scaled_offline.pth`。配置 `config.yaml: cotracker_model_path` 后跑 FiVE eval。 | +| FiVE | Running Time | Per Frame / Running Time | 传统/客观统计 | 按每个生成样本统计；我们 val20=120、val100=600。 | 不直接读视频；读耗时记录和帧数。 | 秒/帧或总秒数；低更好。 | 无需模型；需要记录生成耗时和帧数。 | 正确。论文表格里有 per-frame runtime。这个不是模型质量指标，需要记录生成耗时。 | 不需要模型；需要在生成阶段保存 wall time。 | 对我们的 6 组结果：如果之前没有记录，只能补测同样生成命令的耗时，或标为缺失。建议 manifest 增加 `runtime_sec` 和 `num_frames`，计算 `runtime_sec / num_frames`。 | +| FiVE | FiVE-Acc | FiVE-YN-Acc | VLM-as-judge | 官方 420 pairs；我们可批量 val20=120、val100=600。 | 代码可读 mp4/帧目录/图片；VLM 实际看抽帧/视频输入。 | 0-100% 或 0-1 accuracy；高更好。 | 有 prompt，但代码没有 system role；user prompt 为 `Answer the following question using only YES or NO`；需 Qwen2.5-VL-7B-Instruct。 | 正确。论文和 README 都说明用 Qwen2.5-VL 类 VLM 判断源对象是否消失、目标对象是否出现。 | `Qwen/Qwen2.5-VL-7B-Instruct`；`qwen-vl-utils`；FiVE edit prompts 中的对象字段/问题模板。 | 下载：`hf download Qwen/Qwen2.5-VL-7B-Instruct --local-dir models/Qwen2.5-VL-7B-Instruct`。正式流程：配置 `five_acc_vlm_model_id` 后跑 FiVE eval。我们的数据若没有 source/target object，不能自动生成可靠 YN 问题。 | +| FiVE | FiVE-Acc | FiVE-MC-Acc | VLM-as-judge | 官方 420 pairs；我们可批量 val20=120、val100=600。 | 代码可读 mp4/帧目录/图片；VLM 实际看抽帧/视频输入。 | 0-100% 或 0-1 accuracy；高更好。 | 有 prompt，但代码没有 system role；user prompt 为 `Select the correct answer from the given choices`；需 Qwen2.5-VL-7B-Instruct。 | 正确。多选题判断编辑后识别对象是否为 target object。 | 同 FiVE-YN。 | 同上；必须有 source object、target object、候选项。 | +| FiVE | FiVE-Acc | FiVE-union / FiVE-intersection / FiVE-Acc | VLM-as-judge | 官方 420 pairs；我们可批量 val20=120、val100=600。 | 代码可读 mp4/帧目录/图片；VLM 实际看抽帧/视频输入。 | 0-100% 或 0-1 accuracy；高更好；FiVE-Acc 为多个 accuracy 的平均。 | 由 YN/MC 两类 VLM 判断结果汇总；无独立 system prompt；需 Qwen2.5-VL-7B-Instruct。 | 正确。README 定义 `U-Acc` 为任一问题正确，`∩-Acc` 为全部正确，FiVE-Acc 为这些指标平均。 | 同 FiVE-YN。 | 由 FiVE 官方 `metrics_calculator.py` 汇总。对我们的数据需要先构造 FiVE 风格 QA schema。 | +| OpenVE-3M / OpenVE-Bench | Spatially-Aligned edits | Global Style / Background Change / Local Change / Local Remove / Local Add / Subtitles Edit | VLM-as-judge | 官方 431 rows；CSV 每行一个 edit。 | 直接输入视频：original_video + edited_result_path。 | VLM judge 分数；具体量纲由脚本/API 输出决定；高更好。 | 有类别级长 prompt/rubric。Seed/Doubao 脚本用 `role=system`；Qwen3VL/InternVL 脚本把 rubric 放 user message。需 Seed/Gemini API 或 InternVL3.5-38B/Qwen3VL-32B。 | 基本正确。OpenVE-Bench README 将 benchmark 划分为编辑类型，评价用 Seed1.6VL/Gemini 2.5 Pro/InternVL3.5-38B/Qwen3VL-32B 作为 judge。原表写“用 Seed/Gemini 打分”不完整，因为代码也提供本地 InternVL/Qwen3VL。 | OpenVE-Bench 数据；API key 或本地 VLM。 | 数据下载：`huggingface-cli download --repo-type dataset Lewandofski/OpenVE-Bench --local-dir OpenVE-Bench`。生成结果 CSV：`edited_type,prompt,original_video,edited_result_path`。评测：`python seed_benchmark.py --input_csv new_result.csv --root_path OpenVE-Bench --output_csv out_seed.csv`；或 `python gemini_benchmark.py ...`；或 `python qwen3vl_benchmark.py --model_path models/Qwen3-VL-32B-Instruct ...`。 | +| OpenVE-3M / OpenVE-Bench | Non-Spatially-Aligned edits | Camera Multi-Shot Edit / Creative Edit | VLM-as-judge | 官方 431 rows；CSV 每行一个 edit。 | 直接输入视频：original_video + edited_result_path。 | VLM judge 分数；具体量纲由脚本/API 输出决定；高更好。 | 有类别级长 prompt/rubric。Seed/Doubao 脚本用 `role=system`；Qwen3VL/InternVL 脚本把 rubric 放 user message。需 Seed/Gemini API 或 InternVL3.5-38B/Qwen3VL-32B。 | 正确。属于 OpenVE-Bench 的非空间对齐类编辑，仍用 VLM judge 评分。 | 同上。 | 同上。对我们的现有 val20/100，如果没有 OpenVE 的 `edited_type`，可以用 VLM judge 做自定义评分，但不可称为 OpenVE-Bench 正式分数。 | +| IVE-Bench | Video Quality | Subject Consistency (SC) | 传统/预训练指标 | 官方约 600 videos；JSON 批量逐条评测。 | 需要先转帧目录；README 要 source/target frame folders。 | DINO cosine 连续值；高更好。 | 需 DINO ViT-S/16/相关 DINO backbone。 | 正确。README/代码中质量维度含 `subject_consistency`，实现为 DINO 跨帧相似度。 | DINO；IVEBench 数据和 frame folders。 | 数据下载：`huggingface-cli download --repo-type dataset Coraxor/IVEBench --local-dir IVEBench_DB`。转帧：`python data_process/mp42frames_batch.py --input_path INPUT --output_path FRAMES`。评测：`cd metrics && python evaluate.py --output_path OUT --source_videos_path SRC_FRAMES --target_videos_path EDIT_FRAMES --info_json_path INFO.json --metric subject_consistency`。 | +| IVE-Bench | Video Quality | Background Consistency (BC) | 传统/预训练指标 | 官方约 600 videos；JSON 批量逐条评测。 | 需要先转帧目录；README 要 source/target frame folders。 | CLIP cosine 连续值；高更好。 | 需 CLIP image encoder。 | 正确。代码中为 `background_consistency`，用 CLIP 跨帧相似度。 | CLIP。 | 同 IVE evaluate 命令，把 `--metric background_consistency`。 | +| IVE-Bench | Video Quality | Temporal Flickering (TF) | 传统/客观统计 | 官方约 600 videos；JSON 批量逐条评测。 | 需要先转帧目录；README 要 source/target frame folders。 | 帧间平均差/闪烁强度；低更好。 | 无需模型；代码按帧间 MAE 转为 0-1 flickering score。 | 正确。代码中为 `temporal_flickering`，通过帧间差异衡量闪烁。 | 不需要大模型。 | `cd metrics && python evaluate.py ... --metric temporal_flickering`。 | +| IVE-Bench | Video Quality | Motion Smoothness (MS) | 传统/预训练指标 | 官方约 600 videos；JSON 批量逐条评测。 | 需要先转帧目录；README 要 source/target frame folders。 | 运动先验连续分；高更平滑/更好。 | 需 AMT-G 插帧模型 checkpoint `amt-g.pth`。 | 正确。代码 `path.yml` 指向 AMT-G config/checkpoint，README 说用 video frame interpolation model 的 motion prior。 | AMT-G checkpoint：`amt-g.pth`。 | 下载 AMT-G 到本地后修改 `metrics/path.yml`：`motion_smoothness.checkpoint`。评测：`cd metrics && python evaluate.py ... --metric motion_smoothness`。 | +| IVE-Bench | Video Quality | VTSS | 训练型 reward/evaluator | 官方约 600 videos；JSON 批量逐条评测。 | 需要先转帧目录；README 要 source/target frame folders。 | 训练型连续质量分；高更好。 | 不属于前两类重点；需 `Koala-36M/Training_Suitability_Assessment` 的 `infer.pth`。 | 正确。代码 `path.yml` 指向 `Koala-36M/Training_Suitability_Assessment` 的 `infer.pth`，属于训练型视频训练适任性评分。 | `Koala-36M/Training_Suitability_Assessment` checkpoint。 | 下载：`hf download Koala-36M/Training_Suitability_Assessment --local-dir models/Training_Suitability_Assessment`。修改 `path.yml: vtss.checkpoint`。命令：`python evaluate.py ... --metric vtss`。 | +| IVE-Bench | Instruction Compliance | Overall Semantic Consistency (OSC) | 传统/预训练指标 | 官方约 600 videos；JSON 批量逐条评测。 | 需要先转帧目录；README 要 source/target frame folders。 | VideoCLIP-XL 相似度连续值；高更好。 | 需 `alibaba-pai/VideoCLIP-XL-v2`。 | 正确。代码中为 `overall_semantic_consistency`，用 VideoCLIP-XL-v2 计算目标视频与 target prompt 相似度。 | `alibaba-pai/VideoCLIP-XL-v2/VideoCLIP-XL-v2.bin`。 | 下载：`hf download alibaba-pai/VideoCLIP-XL-v2 --local-dir models/VideoCLIP-XL-v2`。修改 `path.yml`。命令：`python evaluate.py ... --metric overall_semantic_consistency`。 | +| IVE-Bench | Instruction Compliance | Phrase Semantic Consistency (PSC) | 传统/预训练指标 | 官方约 600 videos；JSON 批量逐条评测。 | 需要先转帧目录；README 要 source/target frame folders。 | VideoCLIP-XL 相似度连续值；高更好。 | 需 `alibaba-pai/VideoCLIP-XL-v2`。 | 正确。用 VideoCLIP-XL-v2 计算目标视频与 target phrase 相似度。我们的 val 数据如果没有 `target_phrase`，不能严格复现。 | VideoCLIP-XL-v2；IVE info JSON 中的 `target phrase`。 | 命令：`python evaluate.py ... --metric phrase_semantic_consistency`。 | +| IVE-Bench | Instruction Compliance | Instruction Satisfaction (IS) | VLM-as-judge | 官方约 600 videos；JSON 批量逐条评测。 | 脚本输入帧目录；内部临时转 mp4 给 Qwen。 | Qwen2.5-VL 五分制，1-5；高更好。 | 有 user prompt，无 system role；prompt 要求 Qwen2.5-VL 对 source/target+edit prompt 按 1-5 打分并解释；需 Qwen2.5-VL-72B-Instruct。 | 正确。代码为 `instruction_satisfaction`，用 Qwen2.5-VL 判断目标视频是否执行编辑指令，五分制。 | `Qwen/Qwen2.5-VL-72B-Instruct`，资源很重。 | 下载：`hf download Qwen/Qwen2.5-VL-72B-Instruct --local-dir models/Qwen2.5-VL-72B-Instruct`。修改 `path.yml`。命令：`python evaluate.py ... --metric instruction_satisfaction`。 | +| IVE-Bench | Instruction Compliance | Quantity Accuracy (QA) | 传统/预训练指标 | 官方约 600 videos；JSON 批量逐条评测。 | 需要先转帧目录；README 要 source/target frame folders。 | 0/1 accuracy；高更好。 | 需 GroundingDINO config + `groundingdino_swinb_cogcoor.pth`。 | 正确。代码为 `quantity_accuracy`，使用 GroundingDINO 检测 target span 的数量并和 prompt 目标数量比较。 | GroundingDINO 依赖、config、`groundingdino_swinb_cogcoor.pth`；IVE `target span`。 | 下载：`wget -O models/GroundingDINO/groundingdino_swinb_cogcoor.pth https://github.com/IDEA-Research/GroundingDINO/releases/download/v0.1.0-alpha2/groundingdino_swinb_cogcoor.pth`。安装 GroundingDINO 后命令：`python evaluate.py ... --metric quantity_accuracy`。 | +| IVE-Bench | Fidelity | Semantic Fidelity (SF) | 传统/预训练指标 | 官方约 600 videos；JSON 批量逐条评测。 | 需要先转帧目录；README 要 source/target frame folders。 | VideoCLIP-XL 源/目标视频相似度；高更好。 | 需 `alibaba-pai/VideoCLIP-XL-v2`。 | 正确。代码为 `semantic_fidelity`，用 VideoCLIP-XL-v2 计算源视频与目标视频语义相似度。 | VideoCLIP-XL-v2。 | 命令：`python evaluate.py ... --metric semantic_fidelity`。 | +| IVE-Bench | Fidelity | Motion Fidelity (MF) | 传统/预训练指标 | 官方约 600 videos；JSON 批量逐条评测。 | 需要先转帧目录；README 要 source/target frame folders。 | CoTracker 轨迹保真连续分；高更好。 | 需 CoTracker3 `baseline_offline.pth`。 | 正确。代码为 `motion_fidelity`，用 CoTracker3 提取轨迹而不是传统光流。 | `facebook/cotracker3/baseline_offline.pth`。 | 下载：`hf download facebook/cotracker3 baseline_offline.pth --local-dir models/cotracker3`。修改 `path.yml`。命令：`python evaluate.py ... --metric motion_fidelity`。 | +| IVE-Bench | Fidelity | Content Fidelity (CF) | VLM-as-judge | 官方约 600 videos；JSON 批量逐条评测。 | 脚本输入帧目录；内部临时转 mp4 给 Qwen。 | Qwen2.5-VL 五分制，1-5；高更好。 | 有 user prompt，无 system role；prompt 要求 Qwen2.5-VL 判断除编辑指令外内容保留程度，按 1-5 打分；需 Qwen2.5-VL-72B-Instruct。 | 正确。代码为 `content_fidelity`，用 Qwen2.5-VL-72B 评估未编辑内容是否被保留，五分制。 | `Qwen/Qwen2.5-VL-72B-Instruct`。 | 命令：`python evaluate.py ... --metric content_fidelity`。对我们的现有 val 输出可适配，但需要 source prompt/edit prompt/target video 的规范输入。 | +| VEFX-Bench | Reward Model | Instruction Following (IF) | 训练型 reward/evaluator | 单条或 CSV 批量；benchmark 300 pairs；我们 val20=120、val100=600。 | 直接输入 mp4：original + edited + instruction。 | 连续 1-4 分；高更好。 | 有内部评估 prompt 模板，但推理用 reward head，不是直接 VLM-as-judge；需 `xiangbog/VEFX-Reward-4B`。 | 正确。README 定义 IF 为编辑是否准确反映 instruction；VEFX-Reward-4B 可用。 | `xiangbog/VEFX-Reward-4B`，约 10GB VRAM，bf16。 | 下载：`hf download xiangbog/VEFX-Reward-4B --local-dir models/VEFX-Reward-4B`。安装：`cd reference/benchmarks/edit/code/VEFX-Bench && pip install -r requirements.txt && pip install -e .`。CSV：`original_video,edited_video,instruction`。命令：`python examples/batch_scoring.py --csv edits.csv --output vefx_scores.csv --model models/VEFX-Reward-4B`。 | +| VEFX-Bench | Reward Model | Rendering Quality (RQ) | 训练型 reward/evaluator | 单条或 CSV 批量；benchmark 300 pairs；我们 val20=120、val100=600。 | 直接输入 mp4：original + edited + instruction。 | 连续 1-4 分；高更好。 | 有内部评估 prompt 模板，但推理用 reward head，不是直接 VLM-as-judge；需 `xiangbog/VEFX-Reward-4B`。 | 正确。README 定义 RQ 为 visual clarity、temporal consistency、physical plausibility。 | 同 VEFX IF。 | 同一条 VEFX-Reward 输出中包含 `RQ`。 | +| VEFX-Bench | Reward Model | Edit Exclusivity (EE) | 训练型 reward/evaluator | 单条或 CSV 批量；benchmark 300 pairs；我们 val20=120、val100=600。 | 直接输入 mp4：original + edited + instruction。 | 连续 1-4 分；高更好。 | 有内部评估 prompt 模板，但推理用 reward head，不是直接 VLM-as-judge；需 `xiangbog/VEFX-Reward-4B`。 | 正确。README 定义 EE 为只修改目标区域、避免副作用。 | 同 VEFX IF。 | 同一条 VEFX-Reward 输出中包含 `EE`。这是最适合先评测我们 6 组结果的 benchmark 类 reward。 | +| VEditBench | Semantic Fidelity | Spatial Alignment | 传统/预训练指标 | 未在本地代码固定；按样本逐条/批量评测。 | 通常先抽帧/帧目录评测。 | CLIP 相似度连续值；高更好。 | 需 CLIP。 | 基本正确。论文提取文本显示使用 CLIP 做单帧/空间语义对齐；本地没有官方完整 eval code，需要自实现或找官方 release。 | CLIP。 | 可自实现：抽帧，计算每帧 CLIP image embedding 与 target prompt text embedding cosine，平均。若要正式可比，需要 VEditBench 数据和官方脚本。 | +| VEditBench | Semantic Fidelity | Spatio-Temporal Alignment | 传统/预训练指标 | 未在本地代码固定；按样本逐条/批量评测。 | 直接视频或帧序列，取决于 scorer 实现。 | ViCLIP 相似度连续值；高更好。 | 需 ViCLIP。 | 正确。使用 ViCLIP 衡量整段视频与 target prompt 的时空语义对齐。 | ViCLIP 权重。 | 自实现：用 ViCLIP 对视频和文本编码后算 cosine。无本地官方代码，需另配 ViCLIP repo/model。 | +| VEditBench | Semantic Fidelity | Structural Similarity | 传统/客观统计 | 未在本地代码固定；按样本逐条/批量评测。 | 通常先抽帧/帧目录评测。 | SSIM 0-1；高更好。 | 无需模型；需 SSIM 实现。 | 正确。用 SSIM 衡量源视频与编辑后视频结构保真。 | 不需要大模型。 | 可直接逐帧 SSIM 平均：`python`/`skimage.metrics.structural_similarity`。正式可比需同采样策略、分辨率、帧对齐。 | +| VEditBench | Semantic Fidelity | Motion Similarity | 传统/预训练指标 | 未在本地代码固定；按样本逐条/批量评测。 | 视频或帧序列；运动模型内部抽帧/轨迹。 | CoTracker 轨迹相似度连续值；高更好。 | 需 CoTracker checkpoint。 | 正确。用 CoTracker 抽轨迹并匹配源/编辑视频运动模式。 | CoTracker checkpoint。 | 下载 CoTracker3 后自实现轨迹相似度；可复用 IVE/FiVE 的 CoTracker 环境。 | +| VEditBench | Visual Quality | Image Quality | 传统/预训练指标 | 未在本地代码固定；按样本逐条/批量评测。 | 通常先抽帧/帧目录评测。 | Q-Align image quality 连续分；高更好。 | 需 Q-Align image quality scorer。 | 正确。表格写 Q-Align image quality scorer。 | Q-Align image quality model。 | 需要另下载 Q-Align 相关 scorer；本地 `reference` 没有可直接运行代码。 | +| VEditBench | Visual Quality | Image Aesthetic | 传统/预训练指标 | 未在本地代码固定；按样本逐条/批量评测。 | 通常先抽帧/帧目录评测。 | Q-Align aesthetic 连续分；高更好。 | 需 Q-Align aesthetic scorer。 | 正确。Q-Align aesthetic scorer，训练在 AVA。 | Q-Align aesthetic model。 | 同上。若只要替代指标，可先用 EditBoard LAION aesthetic。 | +| VEditBench | Visual Quality | Motion Smoothness | 传统/预训练指标 | 未在本地代码固定；按样本逐条/批量评测。 | 视频或帧序列；运动模型内部抽帧/轨迹。 | 运动平滑连续分；高更好。 | 需 AMT/VBench motion prior checkpoint。 | 正确。沿用 VBench/插帧模型 motion prior。 | AMT 或 VBench motion smoothness 相关 checkpoint。 | 可复用 IVE-Bench 的 AMT-G 配置做近似实现。正式 VEditBench 需官方实现。 | +| VEditBench | Visual Quality | Temporal Quality | 传统/预训练指标 | 未在本地代码固定；按样本逐条/批量评测。 | 直接视频或帧序列，取决于 scorer 实现。 | Content-Debiased FVD；低更好。 | 需 VideoMAE-v2 / Content-Debiased FVD 实现。 | 正确。Content-Debiased FVD，特征来自 VideoMAE-v2。 | VideoMAE-v2 / CDFVD 实现。 | 本地没有完整代码；需另找 VEditBench 官方实现或自实现 FVD pipeline。 | +| VEditBench | Visual Quality | Video Quality | 传统/预训练指标 | 未在本地代码固定；按样本逐条/批量评测。 | 直接视频或帧序列，取决于 scorer 实现。 | Q-Align video quality 连续分；高更好。 | 需 Q-Align video quality scorer。 | 正确。Q-Align video quality scorer。 | Q-Align video quality model。 | 需要 Q-Align video scorer；当前不建议第一批落地。 | + +## 对我们当前 6 组结果的优先级 + +| 优先级 | 建议 | 原因 | +|---:|---|---| +| 1 | VEFX-Reward-4B | 只需要 source video、edited video、instruction，最容易直接覆盖现有 val20/val100 的 6 组结果。 | +| 2 | EditBoard 无 mask 指标：FF-alpha、FF-beta、CLIP Similarity、Subject/Background Consistency、Aesthetic、Imaging | 能直接适配现有结果；需要先把 mp4 拆帧到 512×512。 | +| 3 | FiVE 的全图 CLIP、NIQE、PSNR/SSIM/LPIPS/MSE、Motion Fidelity | 可部分适配；mask 相关和 FiVE-Acc 需要额外 schema。 | +| 4 | VE-Bench | 如果 checkpoint 能拿到，单条 API 很方便；否则不可复现。 | +| 5 | OpenVE / IVE / VEditBench 全量 | 要正式可比就必须重新跑各自 benchmark 数据，且 IVE/VEditBench 依赖重。 | + +## 我们的 manifest 生成模板 + +```bash +cd /mnt/si002961ale4/default/lgy/shiying/low-high-new + +python3 - <<'PY' +import json +from pathlib import Path + +jobs = [ + ("val20", Path("out/edit_model_face_stage1/eval_outputs"), Path("out/edit_model_face_stage1/eval_samples/val_20.jsonl")), + ("val100", Path("out/edit_model_face_stage1/eval_outputs_val100"), Path("out/edit_model_face_stage1/eval_samples/val_100.jsonl")), +] +methods = ["wan_only", "ditto_global", "full", "text", "vace_hint", "vace_context"] + +for split, root, sample_path in jobs: + out_path = Path(f"out/edit_model_face_stage1/{split}_six_methods_manifest.jsonl") + samples = [json.loads(x) for x in sample_path.read_text(encoding="utf-8").splitlines() if x.strip()] + with out_path.open("w", encoding="utf-8") as f: + for s in samples: + for method in methods: + edited = root / method / f"{s['id']}.mp4" + if not edited.exists(): + raise FileNotFoundError(edited) + row = { + "split": split, + "sample_id": s["id"], + "method": method, + "source_video": s["control_video"], + "edited_video": str(edited), + "instruction": s["prompt"], + "source_prompt": s.get("source_prompt", ""), + "target_prompt": s.get("target_prompt", s["prompt"]), + "mask_path": "", + "target_phrase": "", + "target_span": "", + "category": "", + } + f.write(json.dumps(row, ensure_ascii=False) + "\n") + print(out_path) +PY +```