Infinite Worlds with Versatile Interactions
Zelin Gao, Qiuyu Wang, Jiapeng Zhu, Jingye Chen, Zichen Liu, Qingyan Bai, Jiahao Wang, Yufeng Yuan, Hanlin Wang, Yichong Lu, Ka Leong Cheng, Haojie Zhang, Jian Gao, Tianrui Feng, Yuzheng Liu, Yao Yao, Yinghao Xu, Xing Zhu, Yujun Shen, Hao Ouyang
cs.CV
2026-07-08
LingBot-World-Infinity uses causal pretraining and few-step distillation to achieve drift-free, real-time 720p/60fps interactive world generation for over one continuous hour, with a VLM director and diffusion pilot enabling rich semantic interactions.
Interactive world models face two entangled problems: long-horizon stability and real-time interactivity at high fidelity. In autoregressive generation, each frame is conditioned on frames the model itself produced, so errors compound — textures blur, geometry warps, and scenes drift into incoherence within seconds to minutes. Prior systems routinely sacrifice resolution or frame rate to stay interactive. Beyond the generation backbone, without goal-directed planning a generative model cannot sustain an explorable world across arbitrary time horizons.
The system is trained in two stages and deployed through a layered inference stack.
Causal pretraining formalizes world simulation as a causal generative process: each frame state is conditioned only on past visual observations and the current action input. The central contribution is the MoBA (Mixture of Bidirectional and Autoregressive) attention mask. Standard teacher-forcing masks cause the model to over-rely on the growing context rather than predicting future frames. MoBA appends a fully bidirectional block to the teacher-forcing mask, acting as a regularizer that suppresses this overfitting. A complementary causal cross-attention mask ensures chunk-wise text prompts cannot leak future semantics.
Few-step distillation compresses the multi-step pretrained teacher into a real-time student while suppressing drift over long rollouts. Consistency distillation enforces that points on the same teacher PF-ODE trajectory map to the same prediction; distribution matching distillation (DMD) then refines fidelity by minimizing the KL divergence gradient between the noised student distribution and the noised data distribution. DMD is applied over long self-rollout trajectories, training the student on the state distribution it actually encounters during deployment.
Deployment layers four components: compiler acceleration and hybrid multi-GPU parallelism with async pipelining; a dynamic KV-cache scheduler adapting retained history to the current control signal; a lightweight spatio-temporal refiner for spatial upsampling and frame interpolation; and the Director-Pilot agentic harness pairing a VLM director for high-level semantic planning with a Diffusion Transformer pilot for spatial rendering.
Among all compared systems including closed-source HappyOyster and Genie 3, and open-source Matrix Game 3.0 and DreamWorld, LingBot-World-Infinity is the only model sustaining hour-level continuous generation without visible degradation, while also offering real-time throughput, high dynamic range, rich semantic interaction, and full open-source release.
A single 60-minute uninterrupted session spanning 20 distinct scenarios shows no perceptible quality decay, establishing that stability is structural. The causal pretrained backbone consistently outperforms MAGI-1 and Worldplay under matched settings, retaining sharper textures and stable geometry past the point where competing models degrade.
This work moves the usable time horizon of interactive world models from minutes to hours, demonstrating that pixel-level autoregressive generation and real-time interactivity are not fundamentally incompatible. The 14B teacher plus 1.3B student pairing makes single-GPU deployment accessible. The Director-Pilot framework marks a conceptual shift: the world model ceases to be a stateless frame predictor and begins to behave as a self-sustaining, goal-directed engine — relevant to game generation, embodied simulation, and interactive video interfaces.
Long-term memory is the most fundamental gap. The model stays visually stable but does not truly remember the world it has generated: regions that leave the context window are regenerated rather than recalled, so the world is persistent in appearance but not in identity. Character and style consistency shows subtle drift over very long sessions. Physical plausibility is imperfect because dynamics are learned purely from pixels, occasionally producing intersection artifacts. Real-time operation still requires substantial GPU resources.