Alibaba Open-Sources 24-FPS Character Animation Framework
A character animation system that once needed a skeletal rig now takes the driving video as it is. Alibaba’s Tongyi Wanxiang team has released Wan-Animate-2 as open source, and the framework streams at 24 frames per second while producing results that head-to-head tests found competitive with closed commercial tools.
The change happens before the animation is rendered. Older pipelines extracted a character’s pose from the driving footage, then compressed the motion into features for the model to use. Wan-Animate-2 skips both steps: the raw pixels go directly into a diffusion Transformer, a generative model that learns how the driving performance maps onto the target character. Pandaily reports cleaner hand deformation and fewer limb artifacts on challenging movements.
The framework is not limited to one character or one viewpoint. A single driving video can animate several target characters at once, while the same performance can be rendered from multiple explicit camera positions without retraining. It also handles non-standard aspect ratios and varied character body types, areas where open-source character animation systems have often been brittle.
So what changes in practice? Developers and creators can access locally runnable weights for a task that commercial-grade AI video systems have often left to manual keyframing: making a stylized character perform arbitrary motion while preserving its identity. That could make character animation easier to integrate into downstream video tools, although the release itself does not establish broad production use.
The claim has a clear boundary. The reported comparisons place Wan-Animate-2 alongside ByteDance’s Dreamina and Kuaishou’s KLING MotionControl, but the available account does not describe independent testing or user deployments. For now, the concrete milestone is an open-source framework operating at 24 frames per second, with multi-character and multi-camera capabilities ready for developers to evaluate locally.
Fuentes — leer los originales(hora de París)
Para leer a continuación
- Research robot learns new skills from one video in 29 seconds
- MiniMax H3 opens its weights for unified video-audio workflows
- Omega-0 model reaches 81.8 percent on 11 real home tasks
- X Square Robot's dual-arm gripper robot sorts 1,816 packages in a one-hour livestream; public information says costs cut 70%
Comentarios
Cargando el hilo…
Inicia sesión para escribir un comentario. Iniciar sesión