Many developers building interactive simulations struggle with video models that lose coherence after a few seconds, generate frames slowly, and cannot react to rich user commands in real time. Existing open-source world models either run only for short clips, demand many GPUs, or lack a unified way to steer both high-level semantics and low-level physics. LingBot-World-Infinity addresses these pain points directly.
The model introduces a Mixture of Bidirectional and Autoregressive attention mask (MoBA) that regularizes training and prevents the drift that plagues long-horizon generation. By applying distribution-matching distillation over self-rollout trajectories, the student learns to predict frames that stay consistent for minutes, eliminating the quality decay seen in pure teacher-forcing approaches.
A Director-Pilot harness separates reasoning from rendering. A vision-language model acts as the Director, proposing events such as weather changes or character actions, while a diffusion transformer serves as the Pilot, turning those proposals into pixel-level video at interactive speeds. The public release includes a 14-billion-parameter causal checkpoint that runs on eight GPUs at 480×832 resolution, and a lightweight 1.3-billion variant fits on a single GPU for prototyping.
Users control the simulation with familiar game-style keys: WASD for movement, IJKL for view, Space and P for jump and glide, plus programmable slots for VLM-suggested actions and environmental triggers. Commands are applied at chunk boundaries, guaranteeing that the world state updates predictably without latency spikes.
For those needing higher fidelity, the unreleased deployment stack adds a spatiotemporal refiner that upscales frames and synthesizes intermediate ones, delivering 720p at 60fps on suitable hardware. Because the code is open source under a non-commercial license, teams can experiment, fine-tune, and integrate the model into game prototyping, embodied-AI data generation, or agent evaluation pipelines without building a custom engine from scratch.
In short, LingBot-World-Infinity gives developers a controllable, long-lasting, real-time video world that removes drift, lowers latency, and expands the action space—all in an accessible open-source package.
#AI #MachineLearning #VideoGeneration #Robotics #OpenSource #RealTime