VBench++
VBench++ I2V、long-video 和 trustworthiness 评测,包含仓内 runtime 和 variant-specific 命令。
简介
VBench++ 将 VBench 系列从短 text-to-video 评分扩展到 image-to-video、long-video 和 trustworthiness 维度,覆盖 fairness、bias 和 safety。TPAMI 2025 论文把它定义为一个更通用的视频生成 benchmark suite,用于比较不同生成条件和风险维度下的模型能力。
WorldFoundry 已经把 VBench++ runtime 集成在 worldfoundry/evaluation/tasks/execution/runners/vbench_plus_plus。官方 VBench project 只作为 protocol 参考;WorldFoundry 运行的是仓内 variant runtimes。
官方参考:
- VBench series project: vchitect.github.io/VBench-project
- Paper: arXiv:2411.13503
- VBench-I2V Arena: Vchitect/VBenchI2V_Video_Arena
- 仓内 runner:
worldfoundry/evaluation/tasks/execution/runners/vbench_plus_plus/run_vbench_plus_plus_official_runner.py - 仓内 runtime:
worldfoundry/evaluation/tasks/execution/runners/vbench_plus_plus/runtime
Variants
runner 必须传 --variant。
| Variant | 用途 | Runtime entry |
|---|---|---|
i2v | Image-to-video 评测,输入 reference images 和生成视频。 | runtime/entrypoints/i2v.py |
long | Long-video 评测,关注更长时间范围内的质量和一致性。 | runtime/entrypoints/long.py |
trustworthiness | Culture/fairness、gender bias、skin bias 和 safety。 | runtime/entrypoints/trustworthiness.py |
WorldFoundry 会聚合出:
vbench_plus_plus_i2v_averagevbench_plus_plus_long_averagevbench_plus_plus_trustworthiness_averagevbench_plus_plus_average
资产与生成产物
prompt 和 metadata assets 已经在仓内:
- I2V:
worldfoundry/data/benchmarks/assets/vbench-plus-plus/i2v/vbench2_i2v_full_info.json - Long video:
worldfoundry/data/benchmarks/assets/vbench-plus-plus/long/VBench_full_info.json - Trustworthiness:
worldfoundry/data/benchmarks/assets/vbench-plus-plus/trustworthiness/vbench2_trustworthy.json
候选模型生成视频由用户提供:
export WORLDFOUNDRY_GENERATED_ARTIFACT_DIR=/path/to/vbench-plus-plus/generated_videosI2V custom-image 评测还需要 reference image folder:
export WORLDFOUNDRY_VBENCH_I2V_IMAGE_FOLDER=/path/to/vbench-plus-plus/reference_imagesLong-video custom input 建议把长视频直接放在生成产物根目录下:
/path/to/vbench-plus-plus/generated_videos/
long_video_0001.mp4
long_video_0002.mp4long-video runner 会在需要时准备 upstream metric code 期望的 split-clip layout。
候选模型训练和推理不属于 VBench++ 本身。先用模型 package 生成视频,再导出到对应 variant 需要的 layout。
权重与 Runtime
VBench++ 复用 VBench metric stack,并增加 variant-specific perception/judge assets。先准备 VBench evaluator cache:
export WORLDFOUNDRY_VBENCH_CACHE_DIR=/path/to/cache/models/vbench
export VBENCH_CACHE_DIR="${WORLDFOUNDRY_VBENCH_CACHE_DIR}"常见依赖包括 CLIP/DINO、RAFT、GroundingDINO、SAM-family segmentation、image-quality models、部分 trustworthiness 维度需要的 face detection,以及各 variant scorer 权重。可复用代码在 worldfoundry/base_models,checkpoint 文件仍然是本地运行资产。
运行 I2V
--variant i2v 用于 image-to-video 评分。选中维度需要 aspect ratio 时,通过 --ratio 传入。
cd /path/to/WorldFoundry
PYTHONPATH=. "${WORLDFOUNDRY_UNIFIED_PYTHON:-python}" \
worldfoundry/evaluation/tasks/execution/runners/vbench_plus_plus/run_vbench_plus_plus_official_runner.py \
--variant i2v \
--videos-path "${WORLDFOUNDRY_GENERATED_ARTIFACT_DIR}" \
--full-json-dir "$PWD/worldfoundry/data/benchmarks/assets/vbench-plus-plus/i2v/vbench2_i2v_full_info.json" \
--dimension i2v_subject \
--mode custom_input \
--custom-image-folder "${WORLDFOUNDRY_VBENCH_I2V_IMAGE_FOLDER}" \
--ratio 16-9 \
--output-dir tmp/vbench-plus-plus/i2v_subject \
--json常用 I2V 维度包括 i2v_subject、i2v_background 和 camera_motion。
运行 Long-Video Metrics
--variant long 用于 long-video 评测:
cd /path/to/WorldFoundry
PYTHONPATH=. "${WORLDFOUNDRY_UNIFIED_PYTHON:-python}" \
worldfoundry/evaluation/tasks/execution/runners/vbench_plus_plus/run_vbench_plus_plus_official_runner.py \
--variant long \
--videos-path "${WORLDFOUNDRY_GENERATED_ARTIFACT_DIR}" \
--full-json-dir "$PWD/worldfoundry/data/benchmarks/assets/vbench-plus-plus/long/VBench_full_info.json" \
--dimension temporal_flickering \
--mode long_custom_input \
--output-dir tmp/vbench-plus-plus/long_temporal_flickering \
--timeout 7200 \
--jsonLong-video 维度复用很多 VBench 维度名,例如 subject_consistency、background_consistency、temporal_flickering、motion_smoothness、dynamic_degree、aesthetic_quality、imaging_quality 和 semantic alignment 维度。
运行 Trustworthiness Metrics
--variant trustworthiness 用于 fairness、bias 和 safety:
cd /path/to/WorldFoundry
PYTHONPATH=. "${WORLDFOUNDRY_UNIFIED_PYTHON:-python}" \
worldfoundry/evaluation/tasks/execution/runners/vbench_plus_plus/run_vbench_plus_plus_official_runner.py \
--variant trustworthiness \
--videos-path "${WORLDFOUNDRY_GENERATED_ARTIFACT_DIR}" \
--full-json-dir "$PWD/worldfoundry/data/benchmarks/assets/vbench-plus-plus/trustworthiness/vbench2_trustworthy.json" \
--dimension safety \
--custom-input \
--output-dir tmp/vbench-plus-plus/trustworthiness_safety \
--jsonTrustworthiness 维度包括 culture_fairness、gender_bias、skin_bias 和 safety。Face-related 维度需要 RetinaFace-compatible checkpoint 和 face runtime dependencies。
导入已有结果
导入官方兼容的 VBench++ 结果 JSON 时,需要选择匹配的 variant:
PYTHONPATH=. "${WORLDFOUNDRY_UNIFIED_PYTHON:-python}" \
worldfoundry/evaluation/tasks/execution/runners/vbench_plus_plus/run_vbench_plus_plus_official_runner.py \
--variant long \
--official-results-path /path/to/vbenchpp_eval_results.json \
--videos-path "${WORLDFOUNDRY_GENERATED_ARTIFACT_DIR}" \
--output-dir tmp/vbench-plus-plus/imported_long \
--json输出文件
每次运行会写出:
scorecard.json: WorldFoundry 统一 scorecard,包含 variant dimension scores 和 VBench++ aggregate。raw_metric_table.jsonl: scorecard 使用的 metric rows。dimension_scores.json: dimension-score summary。upstream/*_eval_results.json: 重新计算 metric 时的官方 runtime 输出。upstream_stdout.log和upstream_stderr.log: 仓内 runtime 日志。vbenchpp_long_presplit_manifest.json: 使用long_custom_input时的 long-video split preparation metadata。