LingBot-World-Infinity is an interactive video world model that generates a controllable 720p/60fps world for over an hour without visible drift, by training a Causal video generation backbone with a hybrid attention mask and distilling it into a few-step generator that resists its own error accumulation.
Imagine you’re building a game demo or an embodied-agent playground where a user walks around, swings a sword, and asks for snow, and every frame is generated by a model rather than a game engine. Today’s open interactive world models look great for maybe a minute, then textures smear and geometry warps because each frame is conditioned on prior generated frames and errors compound. Closed systems like Genie 3 hide behind APIs. This paper releases a model that stays stable across a full hour and accepts a real action vocabulary (combat, archery, spell-casting, weather changes), not just camera pans.
The system is a Diffusion Transformer (DiT) that generates video one chunk at a time, conditioned on prior frames plus two action channels: camera pose (encoded as Plücker embeddings) and chunk-wise text prompts. The core training trick addresses a specific failure of pure Teacher Forcing (chunk-wise causal): when every noisy frame can look at a fully clean context, the model learns to lean on that context and stops learning to predict the future, which hurts quality. Their fix, called MoBA (Mixture of Bidirectional and Autoregressive attention), appends a bidirectional block to the teacher-forcing mask so some frames attend to each other freely, acting as a regularizer. A matching cross-attention mask keeps future prompts from leaking into past frames. Training uses a Flow matching objective.
The pretrained model is too slow for interactive use, so it’s compressed with two distillation stages: Consistency distillation to cut denoising steps, then Distribution Matching Distillation (DMD) applied over long self-rollouts (not just teacher-forced states) so the student sees and corrects its own drift.
At deploy time, a Director-Pilot harness wraps the generator. A VLM (Director) watches the current frame, proposes context-appropriate events bound to hotkeys, and the video model (Pilot) renders them.
state = init_frame; kv_cache = []
while user_active:
action = read_input() # WASD, hotkey, or free text
event = director_vlm(state, action) # proposes semantic event
prompt = compose(event, background_prompt)
chunk = generator(state, kv_cache, camera_pose, prompt) # few-step
kv_cache = prune_kv(kv_cache, action) # dynamic scheduling
stream(chunk); state = chunk[-1]
The prevailing recipe for causal video world models is straight teacher forcing plus multi-step diffusion, which produces sharp short clips but degrades within minutes as the rollout feeds its own errors back in. This paper argues the opposite. Long-horizon stability is a training-time property, not a post-hoc filter: mix a bidirectional attention block into teacher forcing so the model doesn’t collapse onto its context, then distill under the model’s own rollout distribution so it learns to correct its own drift. The evidence is the hour-long uninterrupted session with no visible decay, not any single-frame quality score.
The load-bearing result is the single 60-minute continuous generation with no perceptible quality decay across 20 distinct scenarios, which the authors treat as a structural stress test rather than a cherry-picked clip. Prior open causal world models compared against (MAGI-1, others) degrade within seconds to a few minutes on the same kind of rollout.
•
Qualitative comparisons against closed systems Genie 3 and HappyOyster show the distilled model matches or exceeds their visual quality while being the only entry in their comparison table sustaining hour-level generation at real-time throughput in an open release.
•
The action space genuinely extends beyond navigation: character-level combat, archery, spell-casting, shooting, plus environmental interventions like summoning snow or rain, all routed through the Director VLM.
•
The paper reports no quantitative benchmark table with numeric scores. All comparisons are qualitative figures plus a capability-matrix table.
Reach for this when you’re prototyping an interactive demo where a user or agent needs to move through a generated world for longer than a short clip, and you want an open backbone rather than an API. The Director-Pilot pattern is also portable: even outside video, wrapping a generative core with a VLM that proposes hotkey-bound actions is a clean way to give users structured control without hand-scripting a UI.
The authors state they release the models and the agentic harness, and ship both a 14B primary model and a 1.3B lightweight variant that runs on a single GPU. The paper doesn’t provide a repository URL in the text supplied, and doesn’t specify a license. Training data mixes egocentric video, Unreal Engine synthetic footage, and web video from named academic corpora; the training set itself is not released as far as the paper states.
If your autoregressive generator drifts, fix it at training time by mixing a bidirectional block into teacher forcing and distilling on the model’s own rollouts, not on the teacher’s clean states. Post-hoc filters and longer context windows are treating a symptom; the drift is baked in by training exclusively on clean history.
•
The model is visually persistent but not identity-persistent. Leave a region and come back, and it gets regenerated rather than recalled. If your use case depends on returning to the same room with the same objects, this won’t do it.
•
No quantitative benchmark numbers are reported in the paper. Claims of matching or beating Genie 3 and HappyOyster rest on qualitative figures and a capability matrix, so external replication of the ranking isn’t possible from this paper alone.
•
Physics is learned from pixels only, with no explicit geometry or collision model, so characters and objects can intersect or pass through each other. And “real-time” here still means substantial multi-GPU inference, not commodity hardware.