Kling's 3D Move lets you control the camera during motion transfer

If you have spent any time experimenting with AI motion transfer, you know the fatal flaw. You can copy a sick dance routine onto a completely different character, but the camera is practically glued to the floor. The second you try to rotate around the actor or zoom out, the character’s limbs warp into surrealist geometry.

That limitation is finally cracking. On February 3, Kling researchers revealed 3D Move (dubbed 3DiMo in their technical paper), an implicit motion control technique that lets creators transfer human movement while taking full control of the virtual camera.

Ditching the rigid skeleton puppet

Existing motion transfer models, such as OneAnimate or Scale, typically rely on 2D pose estimations or explicit 3D mesh templates like SMPL.

Both approaches come with hefty baggage. 2D poses lock the action strictly to the source camera angle, completely ruling out novel view synthesis. On the other hand, explicit 3D meshes suffer from depth ambiguity and unnatural dynamics. Forcing those rigid external skeletons onto modern video foundation models basically strangles the AI’s native spatial awareness.

3D Move takes a radically different route. Instead of feeding the generator an awkward virtual puppet, it trains a motion encoder alongside the video model. The system distills the source video into compact, view-agnostic motion tokens and injects them via cross-attention.

To give the model real spatial depth, the researchers trained it on a mix of single-view, multi-view, and moving-camera footage. They still used SMPL meshes, but only as a training stabilizer early on, gradually annealing it to zero. In plain English: the model learns to understand genuine 3D movement inside the video world rather than relying on training wheels.

Cinematic camera sweeps and anime characters

Because the motion representation is detached from the camera perspective, users can prompt the camera to zoom, orbit, or pan over the subject without breaking visual consistency.

According to the paper and demonstrated examples, the fidelity holds up across tricky areas like finger articulations and subtle facial expressions. It also generalizes beyond photorealistic humans to stylized anime and 3D animated characters.

In one striking anime demo, as the camera tilts upward, the model contextually renders a blooming cherry blossom branch in the foreground. It demonstrates that the generator is truly reasoning about 3D scene composition rather than simply warping 2D pixels around.

For now, the work is strictly documented in an academic paper on arXiv. Kling has not confirmed when the feature will hit its commercial platform or if weights will be open-sourced.

Why it matters

In AI video generation, dynamic choreography and cinematic camera control have traditionally been mutual enemies. If you wanted dynamic camera moves, you lost consistent character motion; if you wanted accurate human motion, you were stuck with a static lens.

3D Move shows that modern video backbones already possess deep physical and geometric priors if you stop suffocating them with rigid external constraints. By letting the video generator handle the 3D geometry implicitly, Kling has unlocked practical camera direction for motion capture.

For indie filmmakers, animators, and digital artists, this is the missing link between flat clip generation and actual film direction. Instead of scouting for driving footage filmed from the exact right angle, you can take a mundane smartphone recording and turn it into a dynamic orbital shot.

That said, it remains confined to a research paper today. We believe the real test will be inference cost: whether Kling can deliver this level of 3D-aware camera control inside its consumer product without melting server clusters or slowing render queues to a crawl.