Mark the moments, tag the angle, export. Frame-accurate clips, consistent captions, a manifest your trainer reads, and a readiness check before you spend GPU time.
Clipmark is a local desktop app for building motion LoRA datasets from your own videos. You scrub to a moment, mark a 3–6 second clip, pick camera angle and framing from fixed dropdowns, and Clipmark extracts frame-accurate mp4s with consistent captions and a manifest in the format Wan and Hunyuan trainers expect. A readiness check tells you whether the set has enough clips and variety before you rent a GPU. Sold by A2 Digital as a one-time download.
Hand-written ffmpeg commands and caption spreadsheets take a weekend per dataset. Marking in the player takes an hour, and the export does the rest.
Angle and framing come from dropdowns, so every caption uses the same words for the same thing. Consistency is what the LoRA learns.
Clip count, angle coverage and subject variety checked against thresholds you set. No more discovering a thin dataset after the pod bill.
Clipmark loads them in its player with a scrub strip. No upload, no cloud; everything stays on your machine.
Set in and out points for a 3–6 second moment. Pick angle and framing from the vocabulary, add a short action phrase.
Frame-accurate clips cut by ffmpeg at your chosen quality, one caption file per clip, a manifest in musubi-tuner format. Crop and re-extract individual clips as needed.
Readiness report on count and variety. Pro adds one-click rsync to your GPU pod so training starts without a file-transfer step.
Prices in USD. Windows, macOS and Linux. Requires ffmpeg (bundled on Windows and macOS). Optional updates $9/month.
The manifest and caption layout follow musubi-tuner's dataset format, used for Wan 2.x motion LoRAs, and work for Hunyuan Video trainers with the same layout. Other layouts can be added as export profiles.
Yes. Both lists are editable text files. The point is that everyone on a project uses the same one.
No. It builds the dataset and, in Pro, pushes it to your pod. Training runs in your trainer of choice.
No. Clipmark runs locally. The only network action is the optional push to a pod you own.
Anything ffmpeg can read, which is nearly everything. Output is mp4 at the fps and quality you set.
It depends on the motion and the model; the readiness check uses thresholds you set (the defaults reflect common guidance for Wan motion LoRAs) and shows angle coverage so you can see gaps.
Mark them in the player. Export a dataset your trainer can read.
Get Clipmark