Motion is performance. Performance gets paid.

Where our motion data comes from , and where it's going.

This is our public position on the data our models learn from, and on what we owe the people whose movement is in it. It is a versioned document: the version below changes whenever the text does, and every change is recorded at the bottom of the page. Where a statement is not yet backed by something you could check, the page marks it instead of rounding it up.

Last updated: September 1, 2026 · v1.0

Six principles

M—01 / M—06

Each principle is marked for what it is. Practice means we do it now. Commitment means we are undertaking to, and it has not been earned yet.

M—01Motion is performance.The production model we are training under the EU co-funded project Automating the inbetweening of 3D animation using deep-learning models learns only from motion we hold the right to use: performed or authored by people we commissioned and paid for that work, bought under licence from whoever owns it, or released by its owner for free use. Assembling that data is written into the project, not added to it for this page.COMMITMENT

M—02No scraping.We do not train on motion ripped from games, films or video platforms, and we do not buy datasets we cannot audit. Our research prototypes learned from published academic motion-capture sets, named in the model index below with the terms they carry, disclosed datasets, not scraped ones, and not what the production model is being built on.PRACTICE

M—03Your scenes are yours.The motion you send to the hosted backbone is there to return your result. It is not training data for any model we train. Run the backbone yourself and nothing leaves your machine. Our privacy policy says the same, and if that ever changes it becomes an opt-in we have to ask you for.PRACTICE

M—04Receipts, not adjectives.Every model in our product is listed in the model index below: who trained it, what categories of data it learned from, and whether the commitments on this page apply to it. Licences and legal basis are not published per model yet, and the open items are in the register further down. The gaps are printed as gaps rather than rounded up.PRACTICE

M—05Consent is revocable.Contributors to the Motion Library will be able to withdraw a style, and withdrawal only means something if the style then stops being available. Weights do not forget, so this holds only where styles condition generation instead of being baked into the model. That is the architecture the principle requires of us. [TO CONFIRM: architecture decision pending].COMMITMENT

M—06The people who move should get paid twice.Once for the performance, and again each time that performance helps someone animate. The second payment has no mechanism today. Building one is what the Motion Library is for, and until it pays somebody this line stays a commitment rather than a description.COMMITMENT

Model index

I—01 / I—03

The principles above cover the models Animatica trains. Our product also runs a model somebody else trained, so the index says which is which.

I—01

Research prototypes

Trained by

Animatica.

Data sources

Published academic motion-capture sets, AMASS, Human3.6M, Mixamo, Bones, under research and academic terms, several of them non-commercial. Used for prototyping and evaluation, not for the production model.

Our commitments apply?

Yes, we train them, and the terms above are why they stay research work.

I—02

Animatica A0, production model, in training

Trained by

Animatica, under the EU co-funded project.

Data sources

Commissioned performances, paid under contract. The dataset is being assembled now, so this describes its scope and not a finished corpus.

Our commitments apply?

Yes, this is the model M—01 is written for.

I—03

Kimodo, Ardy

Trained by

NVIDIA. We built the workflows around them, not the models.

Data sources

Training-data provenance outside our control. We did not select, collect or licence what these models learned from, and we will not describe it as though we had.

Our commitments apply?

No, our commitments apply to the models we train ourselves.

Generating with either still does not put your motion into a training set; M—03 is about our servers, not about who trained the model.

What we are still cleaning up

C—01 / C—04

A register of open items. Each one closes by being published here.

C—01The historical AMASS breakdownWe can name the collections our prototypes used. We cannot yet publish which sub-dataset went into which run, under which version of its terms. [TO CONFIRM: full list of AMASS sub-datasets used historically]Open

C—02The performer contractWe intend to publish the template we commission motion under, so that what a performer signs is readable before they sign it. [TO CONFIRM: contract template and the date it goes up]Open

C—03The withdrawal architectureM—05 is only as good as the design underneath it, and that design is not settled. [TO CONFIRM: style conditioning confirmed in writing by the ML team]Open

C—04Model names and versionsThe production model has a name now, Animatica A0, but a name is not a version scheme, so we still cannot say which A0 produced a given result. [TO CONFIRM: internal identifiers and a public version scheme]Open

What we are not building

Renunciations

No text-to-finished-film.

We are not building a machine that turns a sentence into a finished shot. We build the frames between your poses, under your direction. The shot stays the animator's to make.

No training on your scenes.

What you send to the hosted backbone comes back to you and goes no further. Self-hosting keeps it off our servers entirely. This one is policy, not aspiration.

No cloning a living artist's style without consent.

A style belongs in the Motion Library because the person who performed it put it there on terms they agreed to. We will not reconstruct someone's motion signature from their published work and sell it.

What this means in practice

Check it yourself

Your keyframes are a constraint.

The model fills the frames between the poses you set. It does not move the poses.

See it in Blender →

The result is editable frames on your rig.

Not a video and not a locked take. Generated motion arrives as keys you can select, retime and delete like your own.

See it in Motion Builder →

You can run it without us.

The generation backbone is open source under Apache-2.0. Self-host it and your scenes never reach our servers.

Hosted or self-hosted →

Motion Library

Collecting interest

M—06 needs a mechanism, and this is the one we are building. Performers and studios contribute style packs, their own motion, on terms they agree to. Those styles condition generation for people who pay to use them, and the contributor earns from that use rather than from a one-time sale. Because a style conditions generation instead of being baked into weights, a contributor who withdraws one actually removes it.

None of that exists yet. There is no marketplace, no payout, and no date we are prepared to publish, announcing one we then miss would cost exactly the credibility this page is built on. What exists is this list, and the intention to open a pilot from it.

What would prove us wrong

Accountability

A page like this is worth something only if it can be shown to be false. Three findings would do it.

Motion in one of our training sets that came from scraping, or from a source we did not licence.

A model of ours trained on the scenes customers sent to the hosted backbone.

A contributor asking us to withdraw a style, and us being unable to remove it.

If you find one, write to hello@animatica.ai. We undertake to answer a question about where our data came from publicly, within 48 hours. A correction gets a line in the changelog below rather than a silent edit.

Changelog

v1.0 , initial publication (September 2026).

Receipts, not adjectives

Every claim above says what backs it, or that nothing does yet.

See the model index Join as a contributor