AI 仓库情报摘要

FR-AI / ANALYSIS

为什么值得关注

该模型通过结合三维缓存实现空间记忆、压缩帧历史保持时间连续性、利用漂移训练和错误库增强稳定性，并采用少步蒸馏实现实时交互，解决了视频世界模型在长时域和可玩性方面的关键挑战。

适合谁使用

视频生成与世界模型的研究人员
寻求交互式长时域环境的AI游戏开发者
探索可控视频合成的创意工作者
研究时空一致性的计算机视觉学者

典型使用场景

通过相机控制交互式探索生成的3D风格视频场景
在生成过程中进行提示词切换的长时域叙事
为游戏环境原型设计实时视频世界模拟
对稳定长视频生成进行基准测试和研究

项目优势

支持实时相机调制和块级提示词切换（每块约1.33秒）的交互能力
利用显式三维缓存和帧历史嵌入实现长时域记忆一致性
通过漂移训练和错误库注入确保分钟级推演稳定性
开源推理代码和预训练权重，可在GitHub和Hugging Face获取

使用前须知

仅开源推理代码，训练代码和训练数据尚未发布
需要依赖外部第三方权重（Gemma文本编码器、Depth-Anything-3），需单独下载且部分需许可
采用非商业许可证（LTX-2社区许可），限制商业用途

README 快速开始

AlayaWorld: Long-Horizon and Playable Video World Generation

Alaya Lab

An interactive autoregressive world model with real-time camera control, prompt switching, and long-horizon memory consistency.

📰 News

[2026-07-21] Full Technical Report released.
[2026-07-16] Inference code released and pretrained weights available on 🤗 Hugging Face. See Quick Start.
[2026-07-08] Project page and technical report released.

🚀 Release Roadmap

Inference code
Pretrained weights — 🤗 AlayaLab/AlayaWorld
Training code
Training data (partial)

✨ Core Properties

AlayaWorld is built around four core properties — interaction, consistency, stability, and runtime.

🎮 Interaction

Two control channels: a rendered 3D cache with lightweight AdaLN camera modulation for grounded, trajectory-aware navigation, and chunk-level prompt switching to introduce new events mid-generation.

🧠 Consistency

Two forms of complementary memory: an explicit 3D cache reprojected to the queried view for spatial recall, plus a compressed frame-history embedding for temporal continuity, so revisited places stay recognizable.

🛡️ Stability

Long-horizon stability from training on drifted histories and an error bank that re-injects accumulated artifacts into both memory and target, preventing errors from compounding over minute-long rollouts.

⚡ Runtime

Real-time interaction via few-step DMD distillation and short temporal chunks, with prompt switching at chunk boundaries to minimize both visual and semantic latency.

🏃 Quick Start

Inference is image-to-video: give the model a first-frame image, a camera trajectory, and a text prompt, and it rolls out a video chunk by chunk along the camera path (1 chunk ≈ 1.33s @ 24fps; ~45 chunks ≈ 1 minute).

1. Environment — a CUDA GPU and PyTorch ≥ 2.6 (the DiT uses flex_attention).

2. Weights — the model runs on four pieces. Only merged_infer.safetensors (AlayaWorld's own weights) is hosted on our 🤗 repo; the text encoder and the depth model are third-party — get them from their original sources:

| Path under checkpoints/ | Source |

项目描述

Full-stack open-source interactive long-horizon world model.

AlayaWorld