WorldFoundry
English中文

AlayaWorld

AlayaWorld: Long-Horizon and Playable Video World Generation

AlayaWorld: image-to-video and camera-controlled-video; Verified.

Edit this pageGitHubMarkdown

AlayaWorld is a camera-controlled interactive world model for long video: from an image it rolls out world video autoregressively with action control and DA3-backed spatial memory. In this catalog it is checkpoint/GPU validated — the released 33.6 GB merged checkpoint passed exact size and SHA-256 validation, and the unified in-tree pipeline generated the official case1 first chunk (32 H.264 frames at 544x960, 24 fps) on an A100 with spatial memory enabled.

Official paper teaser: Interactive world simulation across diverse scenes. AlayaWorld synthesizes explorable worlds that span first- and third-person viewpoints, real-world, game, and synthetic domains, and both indoor and outdoor.
Official paper teaser: Interactive world simulation across diverse scenes. AlayaWorld synthesizes explorable worlds that span first- and third-person viewpoints, real-world, game, and synthetic domains, and both indoor and outdoor.

The recorded paper is AlayaWorld: Long-Horizon and Playable Video World Generation (2026).

Project page: alaya-lab.github.io/AlayaWorld. Code: AlayaLab/AlayaWorld.

The official paper is AlayaWorld: Long-Horizon and Playable Video World Generation (arXiv 2607.06291). Project: alaya-lab.github.io/AlayaWorld. Code: AlayaLab/AlayaWorld. Weights: AlayaLab/AlayaWorld (revision e3f1adeb). License: ltx-2-community-license.

What this repository checked: the released 33.6 GB merged checkpoint passed exact size and SHA-256; transformer / history / Gemma connector keys matched the safetensors header; the official case1 first chunk ran on an A100 with DA3 memory on — 32 H.264 frames at 544×960, 24 fps. Multi-rank Ulysses tensor parity is validated; a true NCCL multi-GPU artifact is pending.

Build your run

Pipeline
AlayaWorldPipelineworldfoundry.pipelines.alayaworld.pipeline_alayaworld
TaskDefault default
Environmentworldfoundry-unified-cu128
DeviceCUDA 12.8
Runnerworldfoundry.pipeline
Loadingfrom_pretrained
Invocationnative_or_standard_invocation
worldfoundry-eval run \
  alayaworld \
  --pipeline.task-profile default \
  --json

What it is for

alayaworld is the recorded camera-control route. Start here to follow a catalog camera path.

The runtime is a thin extension of WorldFoundry's canonical in-tree LTX-2.3 model — Transformer, VAE, Gemma connector, RoPE, patchifier, attention dispatcher, and distributed helpers are all reused; Alaya-specific tensors map only to action AdaLN and the temporal-history encoder. Spatial memory reuses the in-tree Depth Anything 3 model (paper arXiv 2607.06291).

Catalog tasks include image-to-video, camera-controlled-video, interactive-world-model, and long-video-generation. Catalog aliases include alaya-world and alaya_world. Documented modalities: image in, video out.

Recorded evaluation entry points include CameraBench (Camera-control quality of interactive rollouts), MemoBench (Long-horizon spatial-memory consistency evaluation).

Runtime contract

Optional prompt, input_path, frames, fps, and height; artifacts generated_world_video and alayaworld.mp4.

CLI task profiles: default. WorldFoundry binds AlayaWorldPipeline.

Family status: Verified.

Attention dispatch is capability-based; multi-GPU Ulysses context parallelism uses WorldFoundry core collectives. Multi-rank Ulysses tensor parity is validated, but a true NCCL multi-GPU artifact run remains pending multi-GPU availability.

The runtime reuses WorldFoundry's existing LTX-2.3 Transformer, VAE, Gemma connector, RoPE, patchifier, attention dispatcher, and distributed helpers.

Running in WorldFoundry

Launch worldfoundry-eval run alayaworld --pipeline.task-profile default to follow the camera path. The command builder above switches --pipeline.task-profile among recorded tasks and still emits worldfoundry-eval run.

It runs in the unified environment worldfoundry-unified-cu128 (Python 3.11, CUDA 12.8, torch>=2.6).

Recorded weight repositories include AlayaLab/AlayaWorld, google/gemma-3-12b-it-qat-q4_0-unquantized, and depth-anything/DA3NESTED-GIANT-LARGE-1.1. Prepare the environment and check local assets with bash scripts/inference/prepare_model_infer.sh alayaworld. Add --download to fetch missing public checkpoints.

Gated google/gemma-3-12b-it-qat-q4_0-unquantized — accept upstream terms before download; status: Verified.