Astra

Wan2.1-1.3B interactive world with camera actions. Official garden_1 Sekai demo is runner-verified: 832×480, 97 frames, 20 fps. Apache-2.0 bases.
Astra is an interactive world model on a Wan2.1 backbone: one image plus camera actions roll out navigable world video. WorldFoundry vendors the infer-only runtime under worldfoundry/synthesis/visual_generation/kling/astra_runtime and exposes it as AstraPipeline. Studio defaults mirror the official Wan2.1-backed Sekai image rollout.

The official paper is Astra: General Interactive World Model with Autoregressive Denoising (ICLR 2026, arXiv 2512.08931). Project: eternalevan.github.io/Astra-project. Code: EternalEvan/Astra. Weights: EvanEternal/Astra (mirror KwaiVGI/Astra) plus Wan-AI/Wan2.1-T2V-1.3B. Recorded license on those bases: apache-2.0. Publishing institution is not recorded here.
Runner parity is verified: the official garden_1 Sekai demo ran through the in-tree Studio/API path with forward_left camera control and emitted 97 frames at 832×480, 20 fps.
Build your run
AstraPipelineworldfoundry.pipelines.kling.pipeline_astraofficial-sekai-imageworldfoundry.pipelineworldfoundry-eval run \
astra \
--pipeline.task-profile official-sekai-image \
--jsonWhat it is for
Astra world / video route as recorded.
Launch worldfoundry-eval run astra with --pipeline.task-profile official-sekai-image.
Score on CameraBench — camera-following accuracy of interactive rollouts.
Runtime contract
Required: prompt and image; optional: video; actions: camera_direction; artifacts: astra.mp4.
| Knob | Verified demo |
|---|---|
| Size | 832×480 |
| Frames | 97 |
| FPS | 20 |
| Camera | forward_left on garden_1 |
Run in WorldFoundry
Status: Verified. Unified environment worldfoundry-unified-cu128 (Python 3.11, CUDA 12.8). Stage EvanEternal/Astra, Wan-AI/Wan2.1-T2V-1.3B, and EvanEternal/Astra among others locally — missing files fail at load, nothing is downloaded. Run worldfoundry-eval run astra with --pipeline.task-profile official-sekai-image. Catalog VRAM is unset.

