VISTA: Video-Injected Stylized Text-to-Animation

arXiv cs.GR details VISTA, a two-stage framework for generating stylized 3D human motion by combining text-prompted content with style extracted from reference video. The pitch is straightforward but useful for animation pipelines: get expressive motion without needing tightly paired text, video, and stylized motion triplets.

VISTA first uses a dual-channel autoencoder to map motion sequences and video clips into a shared latent space. A masked autoregressive diffusion model then works in that space, with late-fusion style injection through a Dual-AdaLN path so the system can preserve the requested motion structure while borrowing visual style from video.

The training setup is also notable. Instead of relying on paired data, it uses cross-batch unpaired training with latent cycle consistency, letting the model learn from separate semantically rich and stylistically diverse datasets. That makes the approach more practical for studios and tool developers who have lots of motion, lots of video, but not neatly aligned supervision.

In evaluation on rendered mocap references, VISTA reportedly reaches the strongest style recognition accuracy among video-conditioned methods while keeping competitive content alignment. It also exposes decomposed 3-way classifier-free guidance, giving users independent control over the content-style balance at inference, which is the kind of knob animation teams usually want when iterating on character performance.