AlayaWorld：开源可探索视频生成世界

🚀AlayaWorld is an open-source attempt to turn video generation into an explorable world.

1. Control: camera trajectories and chunk-level prompt switching let users navigate and trigger actions while generation is still running.

2. Memory: a 3D cache helps preserve previously visited locations, while compressed frame history keeps recent dynamics in context.

3. Long-horizon runtime: drift-aware training with an Error Bank targets error accumulation; four-step DMD and short chunks aim to keep generation and intent updates responsive.

The key idea is not just “longer video,” but a world that can respond to where you go and what you do.

A natural next question is what happens when an agent enters that world. An agent could plan a goal, navigate, observe the generated feedback, and revise its next action. That would still require a reliable perception layer and real evaluation, but it is an interesting direction for agent-world interaction.

The current results are mainly qualitative, so loop closure, drift, latency, and agent task performance still need broader evaluation. But AlayaWorld is a compelling direction for playable world models.

@alayastd Good Job !! 🥳🥳