You have the model.This is the missing half.

2,400 clips of ASL from 8 Deaf signers, free to download on Hugging Face. Keypoints, gloss annotations with temporal boundaries, and a frozen benchmark: +12 points on a signer the model has never seen, same architecture, same 24 signs. Half of it from fixing the hands, half from the new data.

Three products.One dataset away.

The models already exist, and you can have one running today. What no model can invent is the data underneath it. Each of these three is one model you already have, plus Épée.

Kling 3.0↗ + clercÉpée v0.4 =Avatar

Kling already turns a reference movement into an animated character, hands and all. What it cannot invent is the sign itself. Épée is the dictionary it reads from: every sign cut out of its sentence, timed, signed by real Deaf signers. Pick a word and watch it drive the character.

SIGN · SIGNER

entries724 → 1,501

from 3+ bodies246 → 423

instances3,788 → 10,441

THE SIGN, TIMED

+

YOUR CHARACTER

=

▶

KLING OUTPUT

A RENDER IS NOT A VALIDATION. Motion transfer is approximate, and in a signed language an approximate handshape is a different sign, not a blurrier one. Kling follows the arms and the timing well; the fingers are where it drifts. So choose source clips where both hands stay visible for the whole sign, and have a Deaf signer check the output before anything ships. Épée gives you the movement to drive the model and the timing to cut it exactly; it cannot certify what the model then draws.

HOW THIS ONE WAS MADE. A corpus clip cut at CLERC's annotation boundaries drove Kling 3.0 Motion Control onto one character image. The annotation is what makes the cut exact, and the release ships that timing with the keypoints; the footage itself is never released or licensed. The clips on the right are real renders, watermark included, not mockups.

One fix, twice the data.12 points on a stranger.

Same small BiLSTM, same 24 signs, same held-out Deaf signers, same seeds. Two things changed in the dataset, and we measured them one at a time.

Doubled three times.Same protocol throughout.

Same repo, same format, same pseudonymized signers since v0.1. v0.4 is the newest; every version stays pinned under its git tag and Zenodo DOI. Pick a release.

Asked often.Answered once.