benchmarks/edit/traditional_eval_notes.md

# 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`

优先用 Hugging Face 下载到 `models/`，避免真实机运行时再走 torch hub 联网：

```bash
cd /Users/ouzhang/Desktop/low-high/low-high-new

mkdir -p models/facebook_dino-vitb16 models/facebook_dino-vits16

hf download facebook/dino-vitb16 \
  --local-dir models/facebook_dino-vitb16

hf download facebook/dino-vits16 \
  --local-dir models/facebook_dino-vits16
```

如果跑的 benchmark 原代码仍然写死 `torch.hub.load('facebookresearch/dino:main', ...)`，可以额外在有网机器预热 torch hub 缓存，并把缓存目录放到 `models/torch_cache`：

```bash
cd /Users/ouzhang/Desktop/low-high/low-high-new

mkdir -p models/torch_cache
export TORCH_HOME="$PWD/models/torch_cache"

python3 - <<'PY'
import torch
torch.hub.load('facebookresearch/dino:main', 'dino_vitb16')
torch.hub.load('facebookresearch/dino:main', 'dino_vits16')
PY
```

### LAION Aesthetic Predictor

用途：

- EditBoard `aesthetic_quality`

下载：

```bash
cd /Users/ouzhang/Desktop/low-high/low-high-new

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。`pyiqa` 的权重通常走 torch cache；用 `TORCH_HOME` 可以把下载内容固定在本仓库 `models/torch_cache` 下：

```bash
cd /Users/ouzhang/Desktop/low-high/low-high-new

python3 -m pip install pyiqa

mkdir -p models/torch_cache models/pyiqa
export TORCH_HOME="$PWD/models/torch_cache"
export XDG_CACHE_HOME="$PWD/models/cache"

python3 - <<'PY'
import pyiqa
metric = pyiqa.create_metric('musiq', device='cpu')
print(type(metric).__name__)
PY
```

NIQE 第一次运行会找这个参数文件：

```text
models/torch_cache/hub/pyiqa/niqe_modelparameters.mat
```

如果真实机无法自动下载，可以手动下载到脚本默认读取的位置：

```bash
cd /inspire/hdd/project/intelligentcreativedesign/dangshengqi-253114050252/z-anna/low-high-new

mkdir -p models/torch_cache/hub/pyiqa

hf download chaofengc/IQA-PyTorch-Weights \
  niqe_modelparameters.mat \
  --local-dir models/torch_cache/hub/pyiqa
```

`run_traditional_metrics.py` 默认会设置：

```bash
TORCH_HOME=$PWD/models/torch_cache
HF_HOME=$PWD/models/hf_cache
XDG_CACHE_HOME=$PWD/models/cache
```

因此手动下载到上面的路径后，`--pyiqa-metrics niqe` 会直接从本地读取，不再重新下载。

如果要用 Q-Align/OneAlign 风格的 pyiqa scorer，不建议装进当前训练环境。它需要 `icecream`，并且在 `numpy>=2` 环境里容易触发 torch/动态模块兼容问题。推荐单独建一个 conda 评测环境：

```bash
cd /Users/ouzhang/Desktop/low-high/low-high-new

conda create -n lowhigh-qalign python=3.10 -y
conda activate lowhigh-qalign

# 按机器 CUDA 情况选择 torch 安装方式。A100/CUDA 12.x 可先用 cu121 wheel；
# 如果你的集群有内部 torch 源，也可以替换成集群推荐命令。
python3 -m pip install --upgrade pip
python3 -m pip install torch torchvision --index-url https://download.pytorch.org/whl/cu121

python3 -m pip install \
  git+https://github.com/chaofengc/IQA-PyTorch.git \
  icecream \
  "numpy<2" \
  "transformers>=4.36.1" \
  accelerate \
  sentencepiece \
  einops \
  timm \
  pillow

mkdir -p models/torch_cache models/hf_cache
export TORCH_HOME="$PWD/models/torch_cache"
export HF_HOME="$PWD/models/hf_cache"

python3 - <<'PY'
import pyiqa
metric = pyiqa.create_metric('qalign', device='cpu')
print(type(metric).__name__)
PY
```

说明：

- `pyiqa.create_metric('qalign')` 会 import `pyiqa/archs/q_align/...`，需要额外的 `icecream` 包。
- Q-Align/OneAlign 依赖的部分 torch/动态模块在 `numpy>=2` 环境里可能报 `RuntimeError: Numpy is not available`，所以这里明确 pin `numpy<2`。
- 不建议在 `base` 或训练环境里降级 numpy；优先使用 `lowhigh-qalign` 单独环境。

### LPIPS

用途：

- FiVE background preservation `LPIPS`

安装：

```bash
python3 -m pip install lpips
```

默认 `--lpips-net alex` 会通过 torchvision 下载 AlexNet backbone：

```text
models/torch_cache/hub/checkpoints/alexnet-owt-7be5be79.pth
```

如果真实机无法自动下载，可以手动放到脚本默认读取的位置：

```bash
cd /inspire/hdd/project/intelligentcreativedesign/dangshengqi-253114050252/z-anna/low-high-new

mkdir -p models/torch_cache/hub/checkpoints

wget -O models/torch_cache/hub/checkpoints/alexnet-owt-7be5be79.pth \
  https://download.pytorch.org/models/alexnet-owt-7be5be79.pth
```

如果用 `--lpips-net vgg` 或 `--lpips-net squeeze`，还会分别需要对应 torchvision backbone。当前统一命令默认使用 `alex`。

### CoTracker3

用途：

- FiVE `Motion Fidelity Score`
- IVE `Motion Fidelity`
- VEditBench `Motion Similarity`

下载：

```bash
cd /Users/ouzhang/Desktop/low-high/low-high-new

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
cd /Users/ouzhang/Desktop/low-high/low-high-new

mkdir -p models/VideoCLIP-XL-v2

hf download alibaba-pai/VideoCLIP-XL-v2 \
  --local-dir models/VideoCLIP-XL-v2
```

### GroundingDINO

用途：

- IVE `Quantity Accuracy`

下载：

```bash
cd /Users/ouzhang/Desktop/low-high/low-high-new

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
cd /Users/ouzhang/Desktop/low-high/low-high-new

mkdir -p models/GroundingDINO

if [ -d models/GroundingDINO/code/.git ]; then
  git -C models/GroundingDINO/code pull --ff-only
else
  git clone https://github.com/IDEA-Research/GroundingDINO.git \
    models/GroundingDINO/code
fi
```

如果 GroundingDINO 的依赖和主训练环境冲突，单独建环境：

```bash
cd /Users/ouzhang/Desktop/low-high/low-high-new

conda create -n lowhigh-groundingdino python=3.10 -y
conda activate lowhigh-groundingdino

python3 -m pip install --upgrade pip
python3 -m pip install torch torchvision --index-url https://download.pytorch.org/whl/cu121
python3 -m pip install opencv-python pycocotools matplotlib addict yapf timm
python3 -m pip install -e models/GroundingDINO/code

python3 - <<'PY'
import groundingdino
print("groundingdino ok")
PY
```

### AMT-G

用途：

- IVE / VEditBench `Motion Smoothness`

下载位置按 IVE `metrics/path.yml` 配置到 `amt-g.pth`。可以用 HF repo `lalala125/AMT` 下载到 `models/amt`：

```bash
cd /Users/ouzhang/Desktop/low-high/low-high-new

mkdir -p models/amt

hf download lalala125/AMT \
  --local-dir models/amt

find models/amt -maxdepth 3 -type f | sort
```

下载后把 IVE `metrics/path.yml` 里的 AMT checkpoint 路径指向实际文件，例如：

```bash
cd /Users/ouzhang/Desktop/low-high/low-high-new

find models/amt -name 'amt-g.pth' -o -name '*amt*g*.pth'
```

然后手动编辑 IVE 的 `metrics/path.yml`，把 `motion_smoothness.checkpoint` 改成上面找到的真实路径。注意下面是 YAML 内容，不是 shell 命令，不能直接粘到终端执行：

```yaml
motion_smoothness:
  checkpoint: /Users/ouzhang/Desktop/low-high/low-high-new/models/amt/amt-g.pth
```

如果 AMT/IVE 的 motion smoothness 依赖和主训练环境冲突，单独建环境：

```bash
cd /Users/ouzhang/Desktop/low-high/low-high-new

conda create -n lowhigh-amt python=3.10 -y
conda activate lowhigh-amt

python3 -m pip install --upgrade pip
python3 -m pip install torch torchvision --index-url https://download.pytorch.org/whl/cu121
python3 -m pip install opencv-python imageio imageio-ffmpeg einops tqdm scipy pillow numpy

# 如果使用 IVEBench 的 motion_smoothness 代码，在该环境里进入 IVEBench metrics 目录运行。
# 确保其 path.yml 指向 models/amt 下实际下载到的 amt-g.pth。
```

### Q-Align

用途：

- VEditBench `Image Quality`, `Image Aesthetic`, `Video Quality`

推荐使用 HF model `q-future/one-align`，直接下载到 `models/q-future_one-align`：

```bash
cd /Users/ouzhang/Desktop/low-high/low-high-new

mkdir -p models/q-future_one-align

hf download q-future/one-align \
  --local-dir models/q-future_one-align
```

建议使用上面创建的 `lowhigh-qalign` 环境，避免污染训练环境。如果还没创建，完整命令如下：

```bash
cd /Users/ouzhang/Desktop/low-high/low-high-new

conda create -n lowhigh-qalign python=3.10 -y
conda activate lowhigh-qalign

python3 -m pip install --upgrade pip
python3 -m pip install torch torchvision --index-url https://download.pytorch.org/whl/cu121

python3 -m pip install \
  "numpy<2" \
  "transformers>=4.36.1" \
  accelerate \
  sentencepiece \
  icecream \
  einops \
  timm \
  pillow
```

如果这个新环境里意外装了 CPU 版 `bitsandbytes` 并导致 import 报错，可以直接卸载；OneAlign 推理不依赖它：

```bash
conda activate lowhigh-qalign
python3 -m pip uninstall -y bitsandbytes
```

如果用 Q-Align 官方代码：

```bash
cd /Users/ouzhang/Desktop/low-high/low-high-new

conda activate lowhigh-qalign

mkdir -p models/Q-Align

if [ -d models/Q-Align/code/.git ]; then
  git -C models/Q-Align/code pull --ff-only
else
  git clone https://github.com/Q-Future/Q-Align.git \
    models/Q-Align/code
fi

python3 -m pip install -e models/Q-Align/code
```

快速检查本地 OneAlign 能否加载：

```bash
cd /Users/ouzhang/Desktop/low-high/low-high-new

conda activate lowhigh-qalign

mkdir -p models/hf_cache
export HF_HOME="$PWD/models/hf_cache"
export TRANSFORMERS_CACHE="$PWD/models/hf_cache/transformers"

python3 - <<'PY'
import torch
from transformers import AutoModelForCausalLM

model = AutoModelForCausalLM.from_pretrained(
    "models/q-future_one-align",
    trust_remote_code=True,
    torch_dtype=torch.float16,
    device_map="auto",
)
print(type(model).__name__)
PY
```

如果这里报：

- `ModuleNotFoundError: No module named 'icecream'`：执行 `python3 -m pip install icecream`。
- `RuntimeError: Numpy is not available` 或提示 `A module that was compiled using NumPy 1.x cannot be run in NumPy 2.x`：执行 `python3 -m pip install "numpy<2"`，或者换一个单独的 Q-Align 评测环境。
- `bitsandbytes ... compiled without GPU support`：通常不是 OneAlign 加载的主因；如果它阻断 import，可先卸载 CPU 版 `bitsandbytes` 或在单独环境里重装匹配 CUDA 的版本。

本地 reference 没有完整 VEditBench official eval code。需要另配 Q-Align/OneAlign scorer 后，再接入统一 manifest。

## Run Unified Metrics on val20 / val100

- `build_six_method_manifest.py`: 为 val20/val100 构建统一 manifest。
- `run_traditional_metrics.py`: 默认跑 no-model metrics；按参数额外启用 CLIP、DINO、LAION aesthetic、pyiqa、LPIPS。

如果 manifest 里的 `source_video` 不存在，脚本仍会计算只依赖 edited video 的指标，例如 `clip_t`、`clip_frame_consistency`、`edited_frame_diff_mae`、`temporal_flicker_score`、`laion_aesthetic`、`pyiqa_musiq`、`pyiqa_niqe`。依赖源视频的列会留空，并在 `error` 列记录 `source_load_failed=...`。

需要源视频的列：

- `pixel_mse`
- `pixel_psnr`
- `source_edit_l1`
- `global_ssim`
- `clip_source_edit_similarity`
- `lpips_source_edit`

不需要源视频的列：

- `num_frames_sampled`
- `edited_frame_diff_mae`
- `temporal_flicker_score`
- `clip_t`
- `clip_frame_consistency`
- `dino_frame_consistency`
- `laion_aesthetic`
- `pyiqa_musiq`
- `pyiqa_niqe`
- `pyiqa_qalign_quality`
- `pyiqa_qalign_aesthetic`

### 1. 生成 manifest

`edited_video` 固定来自 `out/edit_model_face_stage1/eval_outputs*`。`source_video` 应该是真实原视频，默认来自 `eval_samples/*.jsonl` 里的 `control_video`；不能用某个 method 的 edited 输出冒充 source，否则 `PSNR/SSIM/LPIPS/source-edit CLIP` 会变成错误比较。

```bash
cd /inspire/hdd/project/intelligentcreativedesign/dangshengqi-253114050252/z-anna/low-high-new

python3 reference/benchmarks/edit/build_six_method_manifest.py \
  --repo-root . \
  --output-dir out/edit_model_face_stage1/traditional_eval_manifests
```

如果真实源视频在另一份目录里，给 builder 加 `--source-root`，脚本会按 `control_video` 的文件名递归搜索并重写 manifest：

```bash
python3 reference/benchmarks/edit/build_six_method_manifest.py \
  --repo-root . \
  --output-dir out/edit_model_face_stage1/traditional_eval_manifests \
  --source-root /path/to/source/videos
```

如果当前机器上的源视频已经被重命名成 hash 文件名，需要提供一个 JSONL 映射文件，至少包含 `sample_id` 和当前视频路径：

```jsonl
{"sample_id": "val_0000", "path": "/inspire/hdd/project/.../datas/ditto_face/low/0a1ec4e5fd6d9079ad1be0ea03cbcfed.mp4"}
```

然后生成 manifest：

```bash
python3 reference/benchmarks/edit/build_six_method_manifest.py \
  --repo-root . \
  --output-dir out/edit_model_face_stage1/traditional_eval_manifests \
  --source-map out/edit_model_face_stage1/source_video_map.jsonl
```

`--source-map` 也支持这些字段名：

- 样本键：`sample_id` 或 `id`
- 当前视频路径：`path`、`video`、`source_video`、`source_path` 或 `control_video`
- 旧路径键：`old_path`、`old_video`、`control_video` 或 `source_video`

注意：如果只有一个目录里一堆 hash mp4，但没有 `sample_id -> hash mp4` 或 `旧文件名 -> hash mp4` 映射，脚本不能安全自动匹配。此时要先从生成数据时的 records/metadata 找回映射。

生成后快速检查路径：

```bash
python3 - <<'PY'
import json
from pathlib import Path

for manifest in [
    Path("out/edit_model_face_stage1/traditional_eval_manifests/val20.jsonl"),
    Path("out/edit_model_face_stage1/traditional_eval_manifests/val100.jsonl"),
]:
    rows = [json.loads(x) for x in manifest.read_text().splitlines() if x.strip()]
    missing_sources = sorted({r["source_video"] for r in rows if not Path(r["source_video"]).exists()})
    missing_edits = [r["edited_video"] for r in rows if not Path(r["edited_video"]).exists()]
    print(manifest, "rows", len(rows), "missing_sources", len(missing_sources), "missing_edits", len(missing_edits))
    for path in missing_sources[:10]:
        print("  missing source:", path)
PY
```

### 2. 主环境一次性跑 CLIP / DINO / LAION / MUSIQ / NIQE / LPIPS

这些指标可以在主环境跑。如果 `pyiqa` 或 `lpips` 没装，先安装：

```bash
python3 -m pip install pyiqa lpips imageio imageio-ffmpeg
```

运行 `val20`：

```bash
cd /inspire/hdd/project/intelligentcreativedesign/dangshengqi-253114050252/z-anna/low-high-new

mkdir -p out/edit_model_face_stage1/traditional_eval_metrics
export TORCH_HOME="$PWD/models/torch_cache"
export HF_HOME="$PWD/models/hf_cache"
export XDG_CACHE_HOME="$PWD/models/cache"

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_full.csv \
  --frames-per-video 16 \
  --resize 256 \
  --device cuda:0 \
  --clip-model-dir models/openai_clip-vit-large-patch14 \
  --dino-model-dir models/facebook_dino-vitb16 \
  --aesthetic-clip-model-dir models/openai_clip-vit-large-patch14 \
  --aesthetic-predictor models/laion_aesthetic/sa_0_4_vit_l_14_linear.pth \
  --pyiqa-metrics musiq,niqe \
  --lpips \
  --clip-batch-size 8 \
  --dino-batch-size 8 \
  --aesthetic-batch-size 8
```

运行 `val100`：

```bash
cd /inspire/hdd/project/intelligentcreativedesign/dangshengqi-253114050252/z-anna/low-high-new

mkdir -p out/edit_model_face_stage1/traditional_eval_metrics
export TORCH_HOME="$PWD/models/torch_cache"
export HF_HOME="$PWD/models/hf_cache"
export XDG_CACHE_HOME="$PWD/models/cache"

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_full.csv \
  --frames-per-video 16 \
  --resize 256 \
  --device cuda:0 \
  --clip-model-dir models/openai_clip-vit-large-patch14 \
  --dino-model-dir models/facebook_dino-vitb16 \
  --aesthetic-clip-model-dir models/openai_clip-vit-large-patch14 \
  --aesthetic-predictor models/laion_aesthetic/sa_0_4_vit_l_14_linear.pth \
  --pyiqa-metrics musiq,niqe \
  --lpips \
  --clip-batch-size 8 \
  --dino-batch-size 8 \
  --aesthetic-batch-size 8
```

输出 CSV 包含：

| Column | 指标来源 / 含义 | 方向 |
|---|---|---|
| `pixel_mse` | FiVE MSE / 全图像素差近似；需要 source + edited | 低更好 |
| `pixel_psnr` | FiVE PSNR 全图近似；需要 source + edited | 高更好 |
| `source_edit_l1` | 源/编辑视频全图 L1 差异；需要 source + edited | 低更好 |
| `global_ssim` | VEditBench / FiVE SSIM 全图近似；需要 source + edited | 高更好 |
| `edited_frame_diff_mae` | IVE temporal flickering 原始帧间差；只需要 edited | 低更好 |
| `temporal_flicker_score` | `(255-frame_diff_mae)/255`；只需要 edited | 高更好 |
| `clip_t` | CLIP edited frame - instruction similarity；只需要 edited + instruction | 高更好 |
| `clip_frame_consistency` | CLIP edited frame cross-frame consistency；只需要 edited | 高更好 |
| `clip_source_edit_similarity` | CLIP source/edit semantic similarity；需要 source + edited | 高更好 |
| `dino_frame_consistency` | DINO edited frame cross-frame consistency；只需要 edited | 高更好 |
| `laion_aesthetic` | LAION aesthetic predictor；只需要 edited | 高更好 |
| `pyiqa_musiq` | MUSIQ image quality, frame average；只需要 edited | 高更好 |
| `pyiqa_niqe` | NIQE no-reference quality, frame average；只需要 edited | 低更好 |
| `lpips_source_edit` | LPIPS source/edit perceptual distance；需要 source + edited | 低更好 |

### 3. 单独环境跑 Q-Align / OneAlign 分数

Q-Align/OneAlign 建议用 `lowhigh-qalign` 环境，避免和主训练环境的 `numpy>=2`、`bitsandbytes` 冲突。先确保环境已创建并能加载 `qalign`：

```bash
cd /inspire/hdd/project/intelligentcreativedesign/dangshengqi-253114050252/z-anna/low-high-new

conda activate lowhigh-qalign

python3 -m pip install imageio imageio-ffmpeg pyiqa
python3 -m pip uninstall -y bitsandbytes || true

export TORCH_HOME="$PWD/models/torch_cache"
export HF_HOME="$PWD/models/hf_cache"
export XDG_CACHE_HOME="$PWD/models/cache"

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_qalign.csv \
  --frames-per-video 16 \
  --resize 256 \
  --device cuda:0 \
  --pyiqa-metrics qalign_quality,qalign_aesthetic

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_qalign.csv \
  --frames-per-video 16 \
  --resize 256 \
  --device cuda:0 \
  --pyiqa-metrics qalign_quality,qalign_aesthetic
```

Q-Align 输出列：

| Column | 含义 | 方向 |
|---|---|---|
| `pyiqa_qalign_quality` | Q-Align / OneAlign image quality scorer | 高更好 |
| `pyiqa_qalign_aesthetic` | Q-Align / OneAlign aesthetic scorer | 高更好 |

### 4. 汇总每个 method 的均值

```bash
cd /inspire/hdd/project/intelligentcreativedesign/dangshengqi-253114050252/z-anna/low-high-new

python3 - <<'PY'
import csv
from collections import defaultdict
from pathlib import Path

paths = [
    Path("out/edit_model_face_stage1/traditional_eval_metrics/val20_metrics_full.csv"),
    Path("out/edit_model_face_stage1/traditional_eval_metrics/val100_metrics_full.csv"),
    Path("out/edit_model_face_stage1/traditional_eval_metrics/val20_metrics_qalign.csv"),
    Path("out/edit_model_face_stage1/traditional_eval_metrics/val100_metrics_qalign.csv"),
]

for path in paths:
    if not path.exists():
        continue
    rows = list(csv.DictReader(path.open()))
    by_method = defaultdict(list)
    for row in rows:
        by_method[row["method"]].append(row)

    print("====", path)
    print("rows", len(rows), "errors", sum(1 for row in rows if row.get("error")))
    fields = [f for f in rows[0].keys() if f not in {"split", "sample_id", "method", "instruction", "source_video", "edited_video", "error"}]
    for method in sorted(by_method):
        print("--", method, "n", len(by_method[method]))
        for field in fields:
            values = []
            for row in by_method[method]:
                value = row.get(field, "")
                if value and value not in {"None", "inf"}:
                    try:
                        values.append(float(value))
                    except ValueError:
                        pass
            if values:
                print(field, sum(values) / len(values))
PY
```

### 5. 为什么 CoTracker / GroundingDINO 不在统一脚本默认跑

- `CoTracker3` 已下载，但 FiVE/IVE 的 motion fidelity 不是简单逐帧 cosine；需要轨迹采样、遮挡处理和 benchmark 自己的 matching/aggregation 逻辑。可以后续单独接一个 `run_cotracker_metrics.py`，不建议混进当前轻量 CSV 脚本。
- `GroundingDINO` 的 Quantity Accuracy 需要每条样本的 `target_span` 和目标数量。当前 `val20/val100` manifest 只有 `instruction/source_video/edited_video`，没有结构化数量字段，所以不能可靠跑 IVE 的 QA。
- `CLIPS.edit`、FiVE background preservation 的正式版本需要 edit mask。当前 manifest 没有 mask，只能跑全图近似 PSNR/LPIPS/MSE/SSIM。