Long-horizon interactive video generation needs two often-competing properties: causal temporal consistency across many frames, and low-latency inference to support live or interactive playback. This project pursues both by combining a causal pretraining paradigm with a distilled "causal-fast" 14B variant, plus an agentic control layer that separates planning (pilot) from environment synthesis (director).
Key Capabilities
- Real-time-capable inference: a distilled 14B "causal-fast" model optimized for chunk-by-chunk KV-cached generation, intended to drive 480–720p streams at interactive frame rates in distilled mode.
- Long-horizon interaction: causal pretraining and chunked frame processing enable stable, multi-hundred-frame sequences with explicit action scripts and events, supporting diverse agent actions (attacking, archery, spell-casting, shooting, etc.).
- Agentic harness: separates a pilot agent (planning character behavior) from a director agent (synthesizing new environment elements), facilitating procedural and open-ended scene progression.
- Practical tooling: includes inference scripts (generate.py), multi-GPU guidance for large-frame counts, and deployment pointers; model weights distributed under CC BY‑NC‑SA 4.0 for non-commercial use.
Who it's for & Tradeoffs
Great fit if you need an off-the-shelf world model for image-to-video research or demos that require long temporal coherence with interactive action scripting, and you have multi‑GPU resources to run 14B inference. Look elsewhere if you need a fully open commercial license, a lightweight single-GPU runtime, or out-of-the-box deployment tooling (the project intentionally omits official deployment code and keeps heavier pretrain checkpoints pending release).
Where It Fits
This project sits between heavyweight pretraining research and real-time demo systems: it aims to bring large-world-model capabilities into interactive settings via distillation. Compared with smaller video diffusion demos, it emphasizes agent-driven scene evolution and long-horizon consistency; compared with end-to-end game engines, it focuses on generative scene synthesis rather than physics-accurate simulation.
Brief Method Notes
Inference uses chunked frame-by-frame generation with KV caching and optional flash-attn acceleration; the repo provides guidelines for multi‑GPU runs, distilled vs. pretrain inference modes, and scripts to reproduce high-frame-count outputs. The team also offers hosted demos (Reactor, LingGuang) to try a subset of features in real time.
