WorldFoundry
English中文

SANA-WM

Method overview: SANA-WM Architecture. Text, video, and pose tokens pass through alternating GDN and softmax attention blocks. Geometry-aware components (UCPE attention and Plücker mixing) are integrated to enable pose-conditioned.

SANA-WM from NVIDIA: world-model and camera-controlled-video; Runner parity pending; dedicated `sana`.

Edit this pageGitHubMarkdown

SANA-WM is a world model built on NVIDIA's efficient Sana architecture: a two-stage design (interactive stage plus LTX-2 refiner) that generates camera-controlled world video with a resident multi-GPU pipeline. A real Workspace run partitioned the components across four A100s and generated eight finite, decodable 1280x704 frames from a nine-frame native window.

Official paper teaser: SANA-WM teaser. From one image and an action trajectory, SANA-WM generates minute-scale 720p worlds with precise control, 64-GPU training, and single-GPU inference.
Official paper teaser: SANA-WM teaser. From one image and an action trajectory, SANA-WM generates minute-scale 720p worlds with precise control, 64-GPU training, and single-GPU inference.
Method overview: SANA-WM Architecture. Text, video, and pose tokens pass through alternating GDN and softmax attention blocks. Geometry-aware components (UCPE attention and Plücker mixing) are integrated to enable pose-conditioned.
Method overview: SANA-WM Architecture. Text, video, and pose tokens pass through alternating GDN and softmax attention blocks. Geometry-aware components (UCPE attention and Plücker mixing) are integrated to enable pose-conditioned.

The recorded paper is SANA-WM: Efficient Minute-Scale World Modeling with Hybrid Linear Diffusion Transformer (2026).

Project page: nvlabs.github.io/Sana/WM. Code: NVlabs/Sana.

The paper is SANA-WM: Efficient Minute-Scale World Modeling with Hybrid Linear Diffusion Transformer. Project page: nvlabs.github.io/Sana/WM. Code: NVlabs/Sana.

Build your run

Pipeline
SanaWMStreamingPipelineworldfoundry.pipelines.sana_wm.pipeline_sana_wm
TaskInteractive Video interactive-video
Environmentsana
DeviceCUDA 12.8
Runnerworldfoundry.pipeline
Loadingfrom_pretrained
Invocationnative_or_standard_invocation
worldfoundry-eval run \
  sana-wm-streaming \
  --pipeline.task-profile interactive-video \
  --json

What it is for

Use sana-wm-streaming for camera-conditioned video inference. The recipe binds the recorded camera-control profile.

Camera actions follow the current official SANA-WM controls: w/s move forward/back, a/d yaw left/right, j/l strafe left/right, and i/k pitch up/down. Held actions use the official default speeds of 0.025 units and 0.6° per frame, with a gradual stop after release. Pass one action for each transition after the initial image, or provide explicit camera_to_world poses.

SANA-WM reuses the canonical Sana network, encoder, autoencoder, scheduler, and recipe roles — no modality-owned Sana runtime remains (paper arXiv 2605.15178, apache-2.0). The interactive pipeline keeps its two-stage model resident and advances with native 8k+1 temporal windows; the quality profile keeps the bidirectional stage and LTX-2 refiner enabled.

Catalog tasks include world-model, camera-controlled-video, and controllable-world-generation. Catalog aliases include sana world model, sana-wm-2.6b, sana-wm-streaming, and Efficient-Large-Model/SANA-WM_streaming.

Recorded evaluation entry points include SANA-WM 80-Scene Benchmark (The dedicated SANA-WM benchmark recorded in this repository), CameraBench (Camera-control adherence of world rollouts).

Not SANA. Score on SANA-WM 80-Scene Benchmark and CameraBench.

Runtime contract

Optional prompt, input_path, interactions, fps, and guidance_scale; artifacts sana-wm-streaming.mp4.

CLI task profiles: interactive-video. WorldFoundry binds SanaWMPipeline.

Family status: Runner parity pending.

The validated four-A100 run is a four-stage smoke that confirms execution and coherent conditioning — it does not claim the default 60-step quality recipe. Component placement across GPUs is configurable.

WorldFoundry registers an in-tree resident pipeline with configurable multi-GPU component placement.

Running in WorldFoundry

Launch worldfoundry-eval run sana-wm-streaming --pipeline.task-profile interactive-video to follow the camera path. The command builder above switches --pipeline.task-profile among recorded tasks and still emits worldfoundry-eval run.

Recorded weight repositories include Efficient-Large-Model/SANA-WM_streaming and Efficient-Large-Model/SANA-WM_bidirectional. Prepare the environment and check local assets with bash scripts/inference/prepare_model_infer.sh sana-wm-streaming. Add --download to fetch missing public checkpoints.

Recorded licenses include apache-2.0; status: Runner parity pending.