Lightricks lets creators live-direct AI videos

LTX Video is an open-source video generation model whose latest release lets you steer content in real time as frames stream in, while extending single-clip generation up to 60 seconds.

Most AI video generators still feel like asynchronous slot machines: you type a prompt, wait several minutes for the cloud to render, and end up with an eight-second clip that might miss the mark. Lightricks updated LTX Video with a new autoregressive video engine that supports live video streaming. The moment the first batch of frames appears on screen, creators can type additional instructions to guide the scene as it unfolds.

What it can do

Live steering turns AI video into an interactive canvas. By refining prompts mid-generation, creators can maintain continuous narratives and adjust camera movements, lighting, or character actions on the fly.

Lightricks co-founder and CTO Yaron Inger noted that the tech shifts AI video from simple prompt-based tricks to an actual storytelling medium where users truly direct the output.

Lightricks outlined several interactive use cases. Game developers could stream in-engine video cutscenes in real time based on player inputs. Augmented reality creators could project AI dancers syncing with live performers at AR concerts. It could also power interactive educational videos that adapt to how students engage with the material.

On the legal side, the model was trained exclusively on licensed stock footage from Getty Images and Shutterstock, keeping generated content clean of copyright infringement worries.

How it runs and specs

The autoregressive engine is baked into two model sizes: the flagship 13-billion parameter (13B) variant and a compact 2-billion parameter (2B) model designed for mobile and edge devices.

Hardware requirements are refreshingly lean compared to proprietary platforms like OpenAI's Sora, Runway Gen-4, or Pika AI 2.1. Lightricks claims the 13B model can output high-resolution video in seconds on a single Nvidia H100 GPU, and can even run with low latency on a standard consumer-grade laptop.

The model weights are completely open on Hugging Face and GitHub for local testing and fine-tuning, with an editing environment available via LTX Studio.

Industry observers see promise alongside technical hurdles. Constellation Research analyst Holger Mueller told SiliconANGLE that streaming video signals a shift toward always-on AI, though he cautioned that the continuous compute required could make operating costs a concern. Furthermore, competitors are branching into other areas: Google's Veo 3 produces its own synchronized audio, while Moonvalley's Marey specializes in motion mimicking.

Why it matters

We think this update hits the two biggest headaches in generative video: clip length and lack of control. Instead of stitching together eight-second snippets in post-production and praying for visual continuity, creators can now produce a full minute of footage and nudge the scene whenever the model drifts off course.

Keeping the weights open and runnable on local hardware makes all the difference. For indie developers, game studios, and researchers looking to integrate generative video into live pipelines without proprietary API lock-in, LTX Video is an easy download worth testing today.