Tora

Tora is a research-driven Video Generation Framework from Alibaba that focuses on steering motion rather than only describing scenes. Instead of relying on text alone, it lets creators specify the paths they want objects to follow, then synthesizes footage that respects those paths while keeping the result visually convincing. The project positions itself as an early system that handles text, visual references, and trajectories together in a single generation pipeline.

What it does

At its core, Tora builds on a Diffusion Transformer backbone. A trajectory extrACTor encodes a supplied motion path into layered spatial-temporal motion tokens, and a motion-guidance fuser injects those tokens into the transformer blocks. A 3D motion variational autoencoder helps preserve movement Information across conSECutive frames. The practical effect is that a user can sketch where something should travel, describe the scene in words, and receive vIDEo where the subject roughly follows that route.

Key capabilities

• Trajectory control: motion paths act as a first-class input, so camera or subject movement can be directed rather than left entirely to chance.
• Multi-condition input: text Prompts, visual content, and trajectories can be combined in one generation request.
• Physical plausibility: the team emphASIzes realistic simulation of movement in the physical world, AIming for high motion fidelity.
• Flexible ouTPUt: the design targets different durations, aspect ratios, and resolutions, and supports extending CLIp length.
• Scalable architecture: the APProach is built to align with the scaling behavior of DiT-style models.

Who it is for

Tora is most relevant to researchers and technical creators exploring controllable video synthesis, as well as studios and animators who need precise stAGIng of movement. Its emphasis on trajectory guidance makes it interesting for previsualization, virtual reality content, and experimental film work where the direction of motion matters. Because the system is presented primarily as a research project, hands-on users should expect a technical workflow rather than a polished consumer editor.

How to get started

Interested users typically begin with the project's official research page, where documentation, dEMOnstrations, and code or model release details are announced. From there, the usual path is to review the paper and Sample outputs, then check whether inference code or weights are available for local or hosted experimentation. Preparing clean text descriptions and simple, well-DeFined motion paths is the most practical way to test the system's behavior.

Pricing and access caveat

No consumer pricing or subscription tiers are described for Tora. Access depends on what the research team releases publicly, and availability of code, weights, or a hosted demo may change over time. Users should verify current licensing and access conditions on the official project page before planning any production use.