Robbyant/lingbot-world

What it solves

LingBot-World is an open-source world simulator designed to generate high-fidelity, interactive video environments. It addresses the challenge of maintaining temporal consistency and robust dynamics across diverse styles (realism, scientific, cartoon) while allowing for real-time user interaction and long-term memory (minute-level horizons).

How it works

Built upon the Wan2.2 framework, the project provides several model versions:

• Base (Cam): Controlled by camera poses (intrinsics and transformation matrices).
• Base (Act): Controlled by specific actions, including a user-friendly action string interface.
• Fast: An accelerated version using KV caching and chunk-by-chunk processing for causal inference, reducing latency to under 1 second for 16 FPS output.

Who it’s for

This tool is intended for developers and researchers in content creation, gaming, and robot learning who need a controllable, interactive world model for simulating environments.

Highlights

• Real-time Interactivity: Achieves sub-second latency for interactive rendering.
• Long-term Memory: Capable of generating consistent videos up to one minute in length.
• Diverse Environments: Supports a wide range of visual styles from realistic to cartoon.
• Controllable Generation: Supports control via camera poses and action strings.
• Quantized Version: Offers a 4-bit quantized model for users with limited GPU resources.