Video to motion
Act it out. Get it on your rig.
Record the motion on your phone, drag two handles to the part you actually want, and get an R15 animation from it. Faster than describing something specific, and far faster than keying it by hand — particularly for timing you can feel but can't write down.
mp4, mov or webm · takes up to 2 minutes · up to 100 MB · one person in frame
Before you upload
What makes a clip work
These checks run on your side before the upload starts and again on the server before any GPU time is spent, so a bad clip costs you nothing but the time to re-shoot it.
Shoot it like this
- One person, fully in frame, head to feet
- Steady camera — a tripod or a propped-up phone
- Even lighting, clear separation from the background
- The action you want somewhere inside the take
These get rejected
- More than one person visible
- Limbs cut off by the frame edge
- Heavy motion blur from a dark room
- The subject small or far from the camera
Rejections name the actual reason — “two people detected in frame”, not “processing failed”. You should never have to guess what to change.
The path
Where your video goes
Trimmed in your browser
Scrub the take, drag the in and out handles to the seconds you want, and preview the selection on a loop. Size and format are checked before a byte is uploaded, so a file that will not work says so immediately rather than after a two-minute upload.
Straight to storage
The file uploads directly to object storage over a presigned URL, and your trim travels with the job as two timestamps. It never passes through our application servers, which is why a 100 MB clip doesn't tie anything up — and why re-running a different five seconds costs no second upload.
Framing checked server-side
Resolution and the number of visible people are probed inside your selected window before the GPU job starts — the walk to the tripod at the start of the take is never looked at. Rejections here cost you nothing.
Deleted on schedule
The inference job reads the file once. Source videos are then deleted on a fixed retention schedule, and you can opt to have them removed immediately after processing.
Timing you can't write down
A hesitation before a swing, a weight shift before a step. These are the things prompts are bad at and a five-second phone clip is perfect for.
Your own performance
Character comes from how a specific person moves. Recording yourself is the shortest path to an NPC that doesn't move like every other NPC.
Nothing kept by default
Uploads exist to be read once by the inference job. Retention is stated plainly in the privacy policy and configurable in settings.
Questions