下载 Checkpoint 与资产

以下命令默认在 WorldAtlas checkout 内执行；若使用共享大盘布局，请确保 WORLDATLAS_CHECKPOINT_ROOT 指向 active checkpoint root。

一键下载

bash scripts/download_metric_checkpoints.sh          # 核心 metric + official runtime 权重
bash scripts/download_metric_checkpoints.sh --all    # + FVMD、JEDi、TRAJAN、HPSv3、TransNetV2、缓存

评测时 checkpoint 缺失会直接 FileNotFoundError — WorldAtlas 不会在打分阶段在线下载权重。

Metric ↔ checkpoint 对照表

指标	ckpt 路径	下载来源
长序列指标、长时域诊断	AMT 模型 amt-s.pth、RAFT 模型 raft-things.pth、MUSIQ SPAQ 权重	HF: lalala125/AMT、RAFT Dropbox、IQA-PyTorch
时间标定、物理帧率	Pulse of Motion 权重	HF: xiangbog/Visual Chronometer
相机误差、重投影误差（DROID）	官方 DROID 权重	Google Drive（gdown）
运动精度	GroundingDINO、SAM、SAM 2.1、BERT base	GroundingDINO、SAM ViT-H、HF: sam2.1-hiera-base-plus、HF: bert-base-uncased
光流误差、运动幅度、运动精度	SEA-RAFT 权重、RAFT Things	HF: Yarimasune/SEA-RAFT，从 RAFT 模型目录复制
运动平滑	VFIMamba 权重	HF: MCG-NJU/VFIMamba（把下载的模型文件重命名为 VFIMamba.pkl）
长序列美学质量	美学预测权重	improved-aesthetic-predictor
重建一致性、met3r 一致性	DA3-GIANT、MASt3R、DINO	HF: DA3-GIANT、HF: MASt3R
背景一致性、记忆回访一致性	CLIP、DINOv2	CLIP、DINOv2
HPSv3	HPSv3 权重、配置、Qwen2-VL	HF: MizzenAI/HPSv3、配置模板、HF: Qwen2-VL-7B-Instruct
片段连续性	TransNetV2 权重	soCzech/TransNetV2
FVMD（视频套件）	PIPs++ 权重	DSL-Lab FVMD PIPs++ 权重
JEDi（视频套件）	JEDi ViT-H 与 SSv2 探针	Meta JEPA
TRAJAN（动态套件）	BootsTAPIR 权重、轨迹自编码器	Google DeepMind BootsTAPIR、TRAJAN autoencoder

In-tree adapter 会在评测时直接从官方指标目录解析文件，具体文件名见下方下载命令。

1. Benchmark metric 权重

CKPT="${WORLDATLAS_CHECKPOINT_ROOT:-ckpt}"

mkdir -p "$CKPT"/{amt_model,raft_model/models,Pulse-of-Motion,musiq,official_metrics/models}

wget -q -O "$CKPT/amt_model/amt-s.pth" \
  https://huggingface.co/lalala125/AMT/resolve/main/amt-s.pth

wget -q -P /tmp https://dl.dropboxusercontent.com/s/4j4z58wuv8o0mfz/models.zip
unzip -o /tmp/models.zip -d "$CKPT/raft_model"

python -m pip install -q huggingface_hub
python - <<'PY'
import os
from huggingface_hub import hf_hub_download
root = os.environ.get("WORLDATLAS_CHECKPOINT_ROOT", "ckpt")
path = hf_hub_download(
    repo_id="xiangbog/Visual_Chronometer",
    filename="vc_common_10_60fps.ckpt",
    local_dir=f"{root}/Pulse-of-Motion",
)
print(path)
PY

wget -q -O "$CKPT/musiq/musiq_spaq_ckpt-358bb6af.pth" \
  https://github.com/chaofengc/IQA-PyTorch/releases/download/v0.1-weights/musiq_spaq_ckpt-358bb6af.pth

2. Official metric 运行时权重

这些权重支撑 motion_accuracy 与 official camera/flow 指标：

CKPT="${WORLDATLAS_CHECKPOINT_ROOT:-ckpt}"
export CKPT

wget -q -P "$CKPT/official_metrics" \
  https://github.com/IDEA-Research/GroundingDINO/releases/download/v0.1.0-alpha/groundingdino_swint_ogc.pth

wget -q -P "$CKPT/official_metrics" \
  https://dl.fbaipublicfiles.com/segment_anything/sam_vit_h_4b8939.pth

wget -q -P "$CKPT/official_metrics" \
  https://huggingface.co/facebook/sam2.1-hiera-base-plus/resolve/main/sam2.1_hiera_base_plus.pt

cp "$CKPT/raft_model/models/raft-things.pth" "$CKPT/official_metrics/models/raft-things.pth"

python -m pip install -q gdown
gdown 1PpqVt1H4maBa_GbPJp4NwxRsd9jk-elh -O "$CKPT/official_metrics/droid.pth"

python - <<'PY'
import os
import shutil
from huggingface_hub import hf_hub_download

root = os.environ["CKPT"]
official = f"{root}/official_metrics"

tartan_src = hf_hub_download(
    repo_id="Yarimasune/SEA-RAFT",
    filename="Tartan-C-T-TSKH-spring540x960-M.pth",
)
shutil.copy2(tartan_src, f"{official}/Tartan-C-T-TSKH-spring540x960-M.pth")

vfimamba_src = hf_hub_download(repo_id="MCG-NJU/VFIMamba", filename="model.pkl")
shutil.copy2(vfimamba_src, f"{official}/VFIMamba.pkl")
print(tartan_src)
print(vfimamba_src)
PY

wget -q -O "$CKPT/official_metrics/sac+logos+ava1-l14-linearMSE.pth" \
  https://github.com/christophschuhmann/improved-aesthetic-predictor/raw/main/sac%2Blogos%2Bava1-l14-linearMSE.pth

if command -v huggingface-cli >/dev/null 2>&1; then
  export HF_HOME="$CKPT/huggingface"
  export HUGGINGFACE_HUB_CACHE="$CKPT/huggingface/hub"
  huggingface-cli download google-bert/bert-base-uncased --local-dir "$CKPT/official_metrics/bert-base-uncased"
else
  python - <<'PY'
from huggingface_hub import snapshot_download
import os
snapshot_download(
    repo_id="google-bert/bert-base-uncased",
    local_dir=os.path.join(os.environ["CKPT"], "official_metrics", "bert-base-uncased"),
)
PY
fi

3. 质量 / 分布类 metric 额外权重

运行 hpsv3_norm、segment_continuity 或 video suite 的 jedi 时需要：

CKPT="${WORLDATLAS_CHECKPOINT_ROOT:-ckpt}"

# HPSv3 奖励模型 + Qwen2-VL backbone
wget -q -O "$CKPT/HPSv3/HPSv3.safetensors" \
  https://huggingface.co/MizzenAI/HPSv3/resolve/main/HPSv3.safetensors
wget -q -O "$CKPT/HPSv3/HPSv3_7B_local.yaml" \
  https://raw.githubusercontent.com/MizzenAI/HPSv3/main/hpsv3/config/HPSv3_7B.yaml
huggingface-cli download Qwen/Qwen2-VL-7B-Instruct --local-dir "$CKPT/Qwen2-VL-7B-Instruct"

# JEDi / V-JEPA（视频分布 metric）
wget -q -O "$CKPT/JEDi/vith16.pth.tar" \
  https://dl.fbaipublicfiles.com/jepa/vith16/vith16.pth.tar
wget -q -O "$CKPT/JEDi/ssv2-probe.pth.tar" \
  https://dl.fbaipublicfiles.com/jepa/vith16/ssv2-probe.pth.tar

# FVMD / PIPs++（视频分布 metric）
wget -q -O "$CKPT/torch/hub/checkpoints/pips2_weights.pth" \
  https://github.com/DSL-Lab/FVMD-frechet-video-motion-distance/releases/download/pips2_weights/pips2_weights.pth

# TRAJAN / BootsTAPIR（单视频 dynamic metric）
wget -q -O "$CKPT/TRAJAN/bootstapir_checkpoint_v2.npy" \
  https://storage.googleapis.com/dm-tapnet/bootstap/bootstapir_checkpoint_v2.npy
wget -q -O "$CKPT/TRAJAN/track_autoencoder_ckpt.npz" \
  https://storage.googleapis.com/dm-tapnet/trajan/track_autoencoder_ckpt.npz

# TransNetV2（segment_continuity）：克隆上游、拉 LFS 权重、转换为 PyTorch
git clone --depth 1 https://github.com/soCzech/TransNetV2 thirdparty/TransNetV2
git -C thirdparty/TransNetV2 lfs pull
cd thirdparty/TransNetV2/inference-pytorch && python convert_weights.py
cp transnetv2-pytorch-weights.pth "$CKPT/TransNetV2/"

或使用仓库脚本：

bash scripts/download_metric_checkpoints.sh --all

4. Hugging Face 与 Torch cache

WorldAtlas 把 HF/Torch cache 固定在 checkpoint root 下，并以 offline 模式运行 bundled adapter。evaluation 前请预加载所需 backend：

CKPT="${WORLDATLAS_CHECKPOINT_ROOT:-ckpt}"

export HF_HOME="$CKPT/huggingface"
export HUGGINGFACE_HUB_CACHE="$CKPT/huggingface/hub"
export TRANSFORMERS_CACHE="$CKPT/transformers"
export TORCH_HOME="$CKPT/torch"

huggingface-cli download openai/clip-vit-base-patch16
huggingface-cli download facebook/dinov2-base
huggingface-cli download depth-anything/DA3-GIANT --local-dir "$CKPT/DA3-GIANT"
huggingface-cli download naver/MASt3R_ViTLarge_BaseDecoder_512_catmlpdpt_metric

python - <<'PY'
import os
import torch
torch.hub.set_dir(os.environ.get("TORCH_HOME", "ckpt/torch"))
torch.hub.load("facebookresearch/dino:main", "dino_deitsmall16")
PY