Smart video-to-image extractor for LoRA training sets, with sane defaults tuned for Anima/SDXL training.
Turn a folder of videos (MMD dance, gameplay, animations) into a clean, deduplicated, properly bucketed image dataset — no coding required.
智能视频转图片训练集工具，专为 LoRA / SDXL 训练优化。
将一个视频文件夹（MMD 舞蹈、游戏录像、动画）转换为干净、去重、分辨率对齐的图片数据集 — 无需编程。
Windows — 不用会编程：
方法 1（最快）： 到 Releases 下载 vid2dataset.exe，双击。
方法 2（zip）： 下载 vid2dataset-*-windows.zip → 右键「全部解压缩」→ 打开文件夹 → 双击 双击启动.bat（或 START.bat）。
然后：选视频文件夹 → 选输出文件夹 → 点提取。
软件右上角有 「检查更新」。
SmartScreen popup? Click More info → Run anyway.
弹出「Windows 已保护你的电脑」？点 更多信息 → 仍要运行。
Full noob guide: HOW_TO_START.txt
- Go to Releases / 前往 Releases
- Download
vid2dataset.exe/ 下载vid2dataset.exe - Double-click to run / 双击运行
No Python needed. / 不需要 Python。
Works on Windows and Linux (NVIDIA). Missing PyTorch never blocks extract.
- Windows .exe: tick GPU 加速 once. The app downloads PyTorch + CUDA 12.6 (12.8 on RTX 50-series) (~2.5-2.7 GB) to
%LOCALAPPDATA%/vid2dataset/gpu_runtime/. Later runs reuse the cache. - Linux (source):
./venv/bin/pip install torch --index-url https://download.pytorch.org/whl/cu126(cu128 for RTX 50). Then tick GPU in the app. - If CUDA is missing, filters stay on CPU and FFmpeg still tries NVDEC/VAAPI.
Windows .exe 首次勾选会下载到本地快取；Linux 需在 venv 里自行安装 CUDA 版 PyTorch。没装也能提取。
Tick Auto-tag images, type your trigger word, extract — every image gets a .txt caption sidecar (trigger, character tags, general tags), ready for kohya / OneTrainer. The first enable downloads the WD tagger model (wd-eva02-large ~1.2 GB or wd-swinv2 ~450 MB) plus onnxruntime (~25 MB) once; inference is GPU-accelerated on any Windows GPU via DirectML, with CPU fallback. / 勾选自动打标并填触发词，每张图会生成可直接训练的 .txt 标签文件。首次启用下载一次模型，任何显卡都可加速。
Already have images? Caption any folder from the CLI:
vid2dataset tag path/to/images --trigger mycharOne-click too blunt? Press Advanced…: scrub each video on a timeline, mark Set In / Set Out segments (the next run extracts only inside them), or step to the exact frame you want and Capture it directly through the same crop/bucket pipeline. / 一键太粗糙？按进阶模式：拖动时间轴、标记片段（下次提取只取片段内的帧），或逐帧找到你要的画面直接截取。
vid2dataset extract videos/ --segment "dance.mp4:30-95.5" --segment "dance.mp4:120-180"git clone https://github.com/peter119lee/vid2dataset.git
cd vid2dataset
pip install -e .[dev] # CPU only
pip install -e .[dev,gpu] # adds torch (CPU build)
pip install torch --index-url https://download.pytorch.org/whl/cu126 --upgrade # CUDA (cu128 for RTX 50-series)Then vid2dataset app to launch the GUI, or vid2dataset extract --help for CLI.
Need help choosing parameters? See USAGE.md for a parameter-by-parameter guide with recommended values for common scenarios.
| OS / 操作系统 | CPU pipeline | GPU pipeline | .exe |
|---|---|---|---|
| Windows 10/11 | ✅ Verified / 已验证 | ✅ NVIDIA CUDA verified | ✅ |
| Linux | ❌ | ||
| macOS Apple Silicon | ❌ | ||
| macOS Intel | ❌ | ❌ |
| GPU vendor / 显卡 | Decode (ffmpeg) | Filter pipeline (PyTorch) |
|---|---|---|
| NVIDIA (any GeForce/RTX/Quadro) | ✅ NVDEC verified | ✅ CUDA verified on RTX 3090 |
| Apple Silicon M1/M2/M3 | ||
| AMD Radeon | ||
| Intel Arc / iGPU | ❌ Not supported by PyTorch |
| Feature | 功能 | Notes |
|---|---|---|
| Scene-aware sampling (PySceneDetect) | 场景感知采样 | Skipped in keyframe mode for speed |
| Blur + brightness quality gate | 模糊 + 亮度质量过滤 | |
| Letterbox auto-crop | 黑边自动裁切 | |
| Bucket-aware resize (Anima/SDXL 64-step grid) | 分辨率桶对齐 | |
| SSIM diversity filter | SSIM 姿态多样性过滤 | GPU-accelerated since v0.4 |
| Color/lighting diversity | 色彩/光照多样性 | |
| Cross-video pHash dedup | 跨视频感知哈希去重 | |
| Auto blur threshold detection | 自动模糊阈值检测 | |
| Per-video minimum guarantee | 每视频最低输出保证 | |
| Subject size filter | 主体大小过滤 | |
| Watermark detection | 浮水印偵測 (v0.6) | warn-only by default |
| Watermark cropping (opt-in) | 浮水印裁切 (v0.7) | |
| Pre-flight HTML report | 訓練前 HTML 報告 (v0.6) | |
| Gallery with hover info | 畫廊滑鼠 hover 資訊 (v0.7) | |
| Auto-tagging (WD tagger) | 自动打标 (v1.0) | .txt sidecars, kohya-ready, DirectML |
| kohya repeats folder | kohya 资料夹结构 (v1.0) | kohya_repeats + flatten |
| Caption quality controls | 标签质量控制 (v1.1) | blacklist / trait pruning / require / exclude |
| Advanced mode: segment cut | 进阶模式：片段选取 (v1.2) | per-video time ranges, GUI + --segment |
| Advanced mode: manual capture | 进阶模式：手动截取 (v1.2) | scrub + capture exact frames |
| Contact sheet + HTML gallery | 缩略图总览 + HTML 画廊 | |
| ETA estimation | 剩余时间预估 | |
| Chinese/English UI | 中文/英文界面切换 | |
| Remember last settings | 记住上次设置 | |
| Cancel mid-extraction | 中途取消 |
| Preset / 预设 | Use case / 用途 |
|---|---|
anima-style |
Style LoRA — broad sampling, auto-quality / 风格 LoRA — 广泛采样 |
anima-character |
Character LoRA — strict quality / 角色 LoRA — 严格质量 |
fast-preview |
Quick preview — JPG @ 768px / 快速预览 |
output/
├── video_one/
│ ├── video_one_00001.png
│ ├── video_one_00002.png
│ └── _stats.json
├── video_two/
│ └── ...
├── _contact_sheet.png
├── _gallery.html <- now with per-image hover info
├── _report.html <- new in v0.6
└── _run_summary.json
Tick Flatten output in the GUI (or flatten_output = true) to drop all images into a single folder for kohya-ss-style trainers.
Real-world benchmark on 34 × 4K MMD videos:
| version | time | speedup |
|---|---|---|
| v0.2.0 | 50.0 min | 1.0× |
| v0.3.1 | 16.0 min | 3.1× |
| v0.4.0 | 14.1 min | 3.5× |
| v0.5.0 | 8.21 min | 6.1× |
| v0.6.0 | 9.31 min | 5.4× (watermark scan) |
| v0.7.0 | ~9 min | ~5.5× |
git clone https://github.com/peter119lee/vid2dataset.git
cd vid2dataset
pip install -e ".[dev]"
pytest # 64+ tests
ruff check src/ tests/
# Build .exe (Windows) — GPU runtime downloads on demand, no separate GPU build
python build_exe.py # vid2dataset.exe (~151 MB)vid2dataset extract path/to/videos -o output --preset anima-style
vid2dataset extract path/to/videos --keyframe --auto-quality
vid2dataset presets
vid2dataset app| Parameter / 参数 | Default / 默认 | Why / 原因 |
|---|---|---|
| resolution | 1024 | SDXL/Anima standard / 标准分辨率 |
| min_pixels | 500,000 | Anima auto-drops below / Anima 自动丢弃 |
| bucket_step | 64 | Anima bucket grid / 桶网格步长 |
| blur_threshold | auto | Auto-detect recommended / 推荐自动检测 |
| phash_distance | 5 | Near-duplicate detection / 近似重复检测 |
| min_per_video | 3 | Never 0 output / 保证不为零 |
vid2dataset prepares training datasets: extraction, curation, captioning. Permanently out of scope: training, upscaling, image editing, prompt tools. / 本工具只做训练集准备（抽帧、筛选、打标）。训练、放大、修图、提示词工具永远不在范围内。
MIT