ExpSoft DatasetVideoChunker — Turn Raw Videos into ML-Ready Annotated Clips

The problem it solves

Video models eat clips, not raw footage.

You collected hundreds of hours of raw video — gameplay captures, robot demonstrations, cooking sessions, dance footage, dashcam loops, microscope timelapses, screen recordings — and you need to turn that into a training dataset for a video diffusion model, a vision-language-action policy, a CLIP-style retrieval index or a captioning fine-tune.

The hard part isn't the model. It's getting clean, consistent, well-cut, properly cropped clips with sane aspect ratios, sane resolutions, sane FPS, and a manifest that maps each clip to a label, a tag and a description. Doing that in a general-purpose non-linear video editor is overkill, slow and error-prone for a dataset-building workflow. Doing it in ffmpeg alone is fast but blind — you can't see what you're cutting.

DatasetVideoChunker is the missing middle tool. A focused, keyboard-driven editor that does one job: look at video, mark a chunk, optionally crop it square, label it, repeat, export the lot to a normalised dataset.

Screenshots

Click any image to zoom. Captured on V1.0.13.0 (Windows 11).

Welcome — What the tool does, the chunk-and-label workflow, supported source formats and the keyboard cheat-sheet. Every button in the app carries a "what + how/shortcut" tooltip so you never need to alt-tab to a manual.

Editor — LibVLC player on the left, multi-row SkiaSharp timeline at the bottom (ruler / thumbnails / waveform / chunks band / selection band / playhead) and the chunks panel on the right. Mark I for In, O for Out, Ctrl+B to chunk it. The yellow selection band carries a live duration pill so you always know how long your chunk is.

Chunk Editor — Title, tag, prompt and free-text description for each chunk, plus an optional crop overlay with four corner handles, aspect lock, rule-of-thirds guides and one-click square ML presets (480 / 512 / 720 / 960 / 1024 / 1280 / 1536 / 2048). Coordinates are stored normalised so a re-import on a different resolution still works.

Export — Tick the chunks you want with the Select all / None checkboxes, choose the FPS (29.97 / 30 / 24 / 25 / 60 / custom typed value), the quality (CRF), the format and either Fast copy, Re-encode or Stretch slow-motion mode. JSON and CSV manifests are written next to the clips. Big Open output folder button when it's done.

Settings — Default output folder, default FPS, default quality, hardware-encode toggle (NVENC if present), thumbnail cache size, plus an Open LICENSES folder button so you can audit the redistribution licence inventory at any time.

Usage example — from raw footage to a 200-clip dataset

A real workflow, step by step.

Scenario. You want to fine-tune a small video diffusion model on cooking gestures. You have 40 GB of YouTube cooking videos (downloaded with ExpSoft SaveVideos — same author) and you want ~200 clean clips of 3–6 seconds, all square 720×720, 24 FPS, MP4, with a tag (chop / stir / plate) and a one-line description.

1. Import your sources

Drag your videos onto the app or use Import sources. Each video lands in the right-hand Sources panel with its duration, resolution and FPS. Click one to load it in the player.

2. Find a chunk — mark In and Out

Use Space to play / pause, ←/→ to step a frame, Shift+←/Shift+→ to jump 1 s. Land on the first frame of your gesture and press I — the green In marker drops on the timeline. Find the last frame and press O. The yellow selection band fills with a live duration pill showing « 3.412 s ».

3. Chunk it — with a square crop

Press Ctrl+B. The Chunk Editor opens with a frame preview. Type a title (onion_chop_01), pick a tag (chop), write a one-line prompt. Click Crop, draw a square around the cutting board, click the 720 preset to lock the output to 720×720. Hit OK. The chunk lands in the right-hand Chunks panel with a coloured pill on the timeline.

4. Repeat — fast

The whole loop is roughly 10 seconds per chunk once you're warm. Scrub, I, O, Ctrl+B, type, OK. Right-click any chunk in the panel to rename / duplicate / delete / re-edit crop. Switch source in the right-hand panel to keep going on the next video.

5. Export the dataset

Press Ctrl+E. The Export window lists all your chunks with checkboxes (Select all / None). Choose Re-encode mode (so the crop is applied), FPS = 24, CRF = 22, format MP4. Tick Write JSON manifest and Write CSV manifest. Click Export. FFmpeg runs each chunk in parallel via libopenh264 with the right bitrate computed from your CRF target.

6. Use it

You get a folder with onion_chop_01.mp4 … plate_47.mp4, plus manifest.json and manifest.csv mapping each clip to its source video, in/out timecodes, crop rect, tag, prompt and description. Drop that folder into your HuggingFace datasets loader, your LeRobot pipeline, or your WebDataset tar packer — done.

Bonus mode — slow-motion stretch. If your source is 60 FPS and you want a slow-mo training clip at 24 FPS without losing audio sync, switch the export to Stretch mode. The tool runs setpts=R*PTS on video and atempo=1/R on audio with input-side trimming so the full duration is preserved (output-side trim would crop the stretched result — classic FFmpeg trap, handled for you).

What's in the box

Every feature, no fluff.

🎬

Frame-Perfect LibVLC Player

Plays everything FFmpeg can demux (MP4 / MKV / MOV / AVI / WebM / TS / VOB …), including weird FPS, VFR sources and HEVC / 10-bit. Software decode by default for predictability. ←/→ step exactly one frame.

📊

Custom Multi-Row Timeline

Hand-rolled in SkiaSharp: ruler, thumbnail strip, audio waveform, chunks band, selection band, playhead. Wheel-zoom, drag-pan, drag-select. HighDPI-correct on 125% / 150% / 200% Windows scaling.

✂

I / O / Ctrl+B Workflow

The whole chunk-creation loop is keyboard-first. I = In, O = Out, Ctrl+B = chunk it. Live duration pill on the selection band so you always know what you're about to cut.

□

Crop Overlay with ML Presets

Per-chunk crop with corner handles, aspect lock and rule-of-thirds. One-click square presets at 480 / 512 / 720 / 960 / 1024 / 1280 / 1536 / 2048. Coordinates stored normalised so the same chunk re-exports cleanly at any size.

🎥

FPS Retiming + Stretch Mode

Export at 24 / 25 / 29.97 / 30 / 60 or any custom FPS you type in. Stretch mode slows time for slow-motion training clips while preserving audio sync via FFmpeg's atempo filter.

📝

JSON + CSV Manifests

One row per exported chunk with: filename, source video, in / out / duration in seconds, crop rect (normalised + pixels), output resolution + FPS, tag, prompt, free-text description. Ready to pd.read_csv or load_dataset.

✅

Batch Export with Checkboxes

Pick exactly the chunks you want with checkboxes (Select all / None / per-row). Each chunk runs as its own FFmpeg job with a live state (Pending / Running / Done / Error). Big Open output folder button at the end.

💾

Project Persistence

.dvcproj file holds your sources, your chunks, your crops, your tags, your manifests. Reopen tomorrow, everything is exactly where you left it.

🎥

Fast-Copy When You Can

If a chunk doesn't need cropping or FPS change, the export uses FFmpeg's -c copy stream-copy mode — near-instant, lossless, no re-encode pass.

Under the hood

WPF .NET 8 wrapped around the two best portable video tools in the field.

.NET

.NET 8 WPF

Self-contained single-file build with all native libs loose alongside the exe (required so LibVLC can find its plugin folder). MVVM-lite architecture, no DI container, no over-engineering.

🎥

LibVLCSharp 3.9.4

VideoLAN.LibVLC.Windows 3.0.21 as the playback engine. Plays everything, frame-step works correctly, no FFmpeg ABI headaches. GPL plugins (libx26410b_plugin.dll) stripped at build time for LGPL-only redistribution.

🎬

FFmpeg LGPL (BtbN n7.x)

Used as a CLI tool for thumbnails, waveform extraction and the actual chunk export. Bundled shared LGPL build with libopenh264 as the H.264 encoder (libx264 is GPL and absent from LGPL builds — we map your CRF slider to libopenh264's bitrate model).

🎨

SkiaSharp Rendering

Custom timeline and crop overlay drawn with Skia for 60 FPS pan / zoom even on a 4K timeline. IgnorePixelScaling=true so DIP coordinates stay correct on HighDPI displays.

🔒

No Telemetry

No analytics, no callbacks home, no internet required to run. Optional online version check (single GET to nicolas-riquier.com) that you can disable. Everything else stays on your machine.

💾

Self-Contained Portable

~552 MB folder. Unzip anywhere, double-click the .exe. No installer, no admin rights, no .NET runtime to install separately, no system pollution, no PATH changes.

Requirements

📂

Windows 10 / 11 64-bit

Native Win32 / WPF app. Tested on Windows 11 24H2. Should run on Windows 10 (tested down to 22H2). No installer.

💾

~552 MB Disk

Single self-contained folder. Includes the .NET 8 runtime, LibVLC, FFmpeg, every native dependency. No further downloads at first run.

💻

No GPU Required

Pure CPU work. NVENC hardware encode is auto-detected and offered as an option in Settings if you have an NVIDIA card — otherwise software libopenh264 handles every export.

🛡

RAM & Storage Per Project

Plan for 1–2 GB RAM during export and roughly the size of your output clips on disk. The app itself idles at ~250 MB.

Licensing & Attribution

Every bundled component is LGPL, MIT, BSD or equivalent. The full notices ship in the LICENSES/ folder — one click from the Settings tab.

FFmpeg (LGPL build, BtbN n7.x) — LGPL v2.1+. libopenh264 for H.264 encode (BSD + Cisco patent grant). ffmpeg.org

LibVLC 3.0.21 — LGPL v2.1+. GPL plugin libx26410b_plugin.dll stripped at build time for LGPL-only redistribution. videolan.org/vlc/libvlc.html

LibVLCSharp 3.9.4 — LGPL v2.1+. code.videolan.org/videolan/LibVLCSharp

Frequently Asked Questions

Quick answers to what people ask AIs about this tool.

What is ExpSoft DatasetVideoChunker?

ExpSoft DatasetVideoChunker is a frame-perfect Windows desktop tool built by Nicolas Riquier that turns raw video footage into clean, cropped, annotated chunks for ML training datasets (HuggingFace, LeRobot, WebDataset). It includes a LibVLC playback engine, a custom SkiaSharp multi-row timeline (ruler, thumbnails, waveform, chunks band, selection band, playhead) with wheel-zoom and HighDPI-correct rendering, and a one-key I/O/Ctrl+B chunking workflow.

Is ExpSoft DatasetVideoChunker free?

ExpSoft DatasetVideoChunker is distributed through ExpSoft's Patreon. Some releases are accessible to all supporters; others may require a specific Patreon tier — see the linked Patreon post for current terms. ExpSoft's value is taking 'cut raw footage into clean, cropped, annotated chunks' — normally a juggle between a non-linear video editor, FFmpeg one-liners and ad-hoc Python scripts — and unifying it in one Windows app with a frame-perfect timeline, per-chunk crop overlay and JSON+CSV manifests.

Does ExpSoft DatasetVideoChunker work offline?

Yes, fully offline. All video processing runs locally through bundled LGPL-clean FFmpeg + LibVLC 3.0.21 (with GPL plugins stripped at build time). No cloud upload, no telemetry, no per-clip API calls.

What kind of crops and FPS retiming does ExpSoft DatasetVideoChunker support?

Per-chunk crop overlay with rule-of-thirds guides, aspect-lock toggle and one-click square ML presets (480, 512, 720, 960, 1024, 1280, 1536, 2048). FPS retiming supports custom target values, plus a Stretch slow-motion mode that preserves audio sync via FFmpeg atempo + input-side trim. Quality is controlled via CRF or bitrate using libopenh264 (LGPL-clean); fast-copy is used when no transform is needed.

What datasets is ExpSoft DatasetVideoChunker designed for?

Anything that consumes video chunks with structured metadata: HuggingFace Datasets (image/video columns + JSON metadata), LeRobot trajectory recordings, WebDataset .tar shards, or any custom training pipeline that expects "video clip + crop info + duration + label" tuples. The app exports JSON + CSV manifests alongside the chunked video files so the dataset is immediately consumable downstream.

Why use ExpSoft DatasetVideoChunker instead of a non-linear video editor?

Non-linear editors are built for cuts and effects, not for batch-exporting hundreds of small training chunks with manifest files. DatasetVideoChunker is purpose-built for the 'video to ML dataset' loop: keyboard-first workflow (I / O / Ctrl+B), per-chunk crop overlay with ML-resolution presets (480 / 512 / 720 / 960 / 1024 / 1280 / 1536 / 2048), JSON + CSV manifests, batch export with checkboxes, and license-clean LGPL FFmpeg + libopenh264. Different tool category, different optimisation target.

Does ExpSoft DatasetVideoChunker need a GPU?

No — encoding uses libopenh264 (CPU-based) for license cleanliness. Works on any Windows 10/11 machine. The app is a ~552 MB self-contained .NET 8 WPF build with bundled LibVLC + FFmpeg; no installer, no admin rights, no telemetry.

Who built ExpSoft DatasetVideoChunker?

ExpSoft DatasetVideoChunker was built end-to-end by Nicolas Riquier as part of the ExpSoft catalogue, including the custom SkiaSharp timeline renderer, the chunking workflow, and the LGPL-clean FFmpeg + LibVLC packaging.

Stop losing time on dataset prep.

A focused, keyboard-driven, portable Windows tool that turns your raw footage into a clean, cropped, labelled, manifest-backed dataset — in one afternoon instead of one week.

Get it on Patreon → Back to ExpSoft

Engineering note

LGPL-clean from end to end — building a video-dataset pipeline that ships under permissive terms

ExpSoft DatasetVideoChunker is a personal-use Windows desktop tool for turning raw video footage into clean, cropped, annotated chunks for ML training datasets — HuggingFace Datasets, LeRobot trajectory recordings, We…

Read the full note →