DeepMotion

There are two companies called DeepMotion and search results mix them constantly.

One is a Beijing autonomous-driving startup that Xiaomi acquired in 2021 to build out its electric-vehicle programme. It has nothing to do with animation.

The other — this one — is DeepMotion, Inc., an independent company that has been working on physics simulation and AI motion capture since 2014. It is not owned by Xiaomi, it has not been acquired, and it makes two products: Animate 3D, which turns video into 3D animation, and SayMotion, which turns text prompts into 3D animation. Its published partner list includes Nvidia, Intel, Qualcomm, Samsung, Epic Games, Bandai Namco, Cornell and MIT.

If you landed here from a search that mentioned Xiaomi, you have the wrong company.

You upload a video of a person moving. You get back a rigged 3D animation of that motion, ready to export into a game engine.

That is the entire pitch, and it is a big one, because the conventional route to that output involves a motion-capture volume, a suit, markers, calibration, and a day rate that indie studios do not have. Animate 3D replaces all of it with a phone camera and a web upload.

Input accepts MP4, MOV and AVI. Output comes back as FBX, BVH, MP4 or GLB — GLB restricted to custom characters. FBX and BVH are the formats that matter: they drop into Unity, Unreal, Blender and Maya without conversion.

The capture supports single-person and multi-person jobs in the same clip, plus a pose mode that extracts a single 3D pose rather than a motion sequence. Face tracking and hand tracking are optional layers on top of body capture, priced separately in credits.

Characters are handled either by DeepMotion's default set, by uploading your own FBX or GLB rig, or by generating one through the company's partnership with Wolf3D.

The second product inverts the input. SayMotion generates 3D animation from a text or video prompt rather than requiring reference video, and it sits on its own pricing ladder and its own API.

The practical use is filling gaps. Animate 3D needs someone to physically perform the motion, which is fine for a walk cycle and impractical for a backflip. SayMotion covers the motions nobody on your team can do in front of a camera. There is also a Motion-to-Video capability in beta.

DeepMotion's pricing unit is the credit, and its definition is unusually clean: one credit equals one second of animation, or one 3D pose.

The complication is the modifiers. Face tracking costs an extra 0.5 credit per second. Hand tracking costs another 0.5 credit per second. Turn both on and a 30-second take costs 60 credits rather than 30 — the price doubles.

Credits reset at the end of each plan month and do not carry over. This is true on annual plans as well as monthly ones, which is worth internalising: paying yearly buys you a discount, not a bank. Your usable capacity is the monthly allotment, every month, whether or not you shot anything.

Running out means waiting for the next cycle or upgrading, and upgrades are prorated with the new tier's credits made available immediately.

The Studio tier removes the cap but not the constraint. Its unlimited credits process at the highest job priority for the first 7,200 credits in a month — two hours of animation — after which further jobs run at low priority. Unlimited in volume, rationed in speed.

The paid tier prices themselves are rendered dynamically on DeepMotion's pricing pages and should be read there before you commit; the structure above is stable, the dollar figures are not something to take from a review.

DeepMotion runs a Free Animation Credit Program that lets you earn back the credits a job consumed. There are two routes.

The first is Animation Labeling: after uploading a video, you write a detailed description of the motion — up to 256 characters, covering action type and style — and a proprietary model scores its quality. Better, longer, more specific descriptions earn more credits back. Specialist vocabulary is encouraged, and DeepMotion's own guidance lists examples across dance, boxing, MMA, basketball and gymnastics.

The second is Animation Correction: open the finished job in the Rotoscope Pose Editor, fix joints that tracked badly, and earn credits based on how much closer your corrected pose sits to a mechanical model of a human skeleton.

DeepMotion is open that this feeds its models — the company states the programme improves both Animate 3D and SayMotion, and that participant text trains its understanding of motion descriptions. Between the two routes you can in principle recover the full credit cost of any job.

That is a reasonable deal, stated honestly by the vendor. It is also unpaid data labelling, and the credits you earn reset monthly like any others. Decide whether your time is worth more than the credits before you spend an evening in the pose editor.

The honest limits are structural to the technique, not to this implementation.

Gross body motion — the trajectory of limbs and torso through space — is what monocular capture does well, and it is what most game animation needs. Walk cycles, combat, dance, idle gestures, sports motion: all reasonable targets.

The hard parts are the ones you would expect. Fine finger articulation. Foot contact with the ground plane, which is where sliding artefacts come from. Limbs occluded by the body during a turn. Depth ambiguity in a single camera view.

DeepMotion's response is twofold: published capture guidelines covering framing, resolution, frame rate and lighting, and the Rotoscope Pose Editor for frame-by-frame cleanup afterwards. The existence of both is the useful signal. This is a tool that expects input discipline and output cleanup — which is also true of real motion capture, just with cheaper equipment.

Support guidance asks users to supply the original video, the output animation and a screenshot of job settings when something looks wrong, which tells you where the failures usually originate.

The Freemium account is genuinely free — no credit card, no expiry, 60 seconds of animation a month. It is also licensed for personal, non-commercial use only.

DeepMotion is specific that commercial rights follow your plan type, not the source of your credits. Animations produced with free earned credits on a paid plan are commercially licensed; animations produced on the Freemium tier are not, no matter how you obtained them. If your project will ship, budget for a subscription from the start rather than discovering the restriction after building a library.

Uploads are private to your account and not visible to other users.

Indie game developers are the obvious fit. A studio that cannot justify a mocap stage but needs fifty character animations can shoot them on a phone in an afternoon and export FBX the same day. The cleanup cost is real but it is measured in hours, not in day rates.

Previz and animatics are the second fit: motion that only has to communicate blocking, where fidelity is irrelevant and speed is everything.

VTubers, avatar work and virtual production round it out, with the Wolf3D character pipeline making the avatar half straightforward.

What this is not is a video generator. If what you want is a finished rendered clip, Runway, Kling and Luma are in a different business — they produce pixels, DeepMotion produces animation data you still have to render. Confusing the two is the fastest way to buy the wrong subscription.