WorldFoundry
English中文

MAGI-1

Method overview: Model Architecture of Transformer-based VAE.

MAGI-1 from Sand AI: image-to-video and video-generation; Runner parity pending; dedicated `magi-official-cu124`.

Edit this pageGitHubMarkdown

MAGI-1 is Sand AI's open autoregressive video generation model, cataloged for image-to-video and general video generation and tracked as a high-priority open candidate with Apache-2.0 HF metadata. The WorldFoundry route is complete and reports missing checkpoint assets at runtime.

Official paper teaser: (Left) Magi-1 performs chunk-wise autoregressive denoising. The video is generated in chunks of 24 frames, where each chunk attends to all previously denoised chunks. Once a chunk reaches a certain denoising level,.
Official paper teaser: (Left) Magi-1 performs chunk-wise autoregressive denoising. The video is generated in chunks of 24 frames, where each chunk attends to all previously denoised chunks. Once a chunk reaches a certain denoising level,.
Method overview: Model Architecture of Transformer-based VAE.
Method overview: Model Architecture of Transformer-based VAE.

The recorded paper is arXiv 2505.13211.

Code: SandAI-org/MAGI-1.

Build your run

Pipeline
MAGI1Pipelineworldfoundry.pipelines.video_official.pipeline_official_video
TaskInteractive Video interactive-video
Environmentmagi-official-cu124
DeviceCUDA 12.4
Runnerworldfoundry.pipeline
Loadingfrom_pretrained
Invocationnative_or_standard_invocation
worldfoundry-eval run \
  magi-1 \
  --pipeline.task-profile interactive-video \
  --json

What it is for

magi-1 is the recorded text/image-to-video line. Use it when you want both prompt and still conditioning on the same catalog card.

Upstream describes MAGI-1 as a large autoregressive video model generating chunk-by-chunk with causal attention, enabling streaming-style continuation. This repository records only the prompt+image input contract and the generated_video artifact; scale and internals are not recorded here.

Catalog tasks include image-to-video and video-generation. Documented modalities: image in, video out.

Recorded evaluation entry points include VBench (Baseline video generation quality suite).

Runtime contract

Optional prompt, negative_prompt, input_path, frames, and fps; artifacts magi-1.mp4.

CLI task profiles: interactive-video. WorldFoundry binds MAGI1Pipeline.

Family status: Runner parity pending.

No runner-parity record exists yet — treat GPU validation as not done. Stage the MAGI-1 weights first; the route fails informatively when they are missing.

Added as a high-priority open video generation candidate with Apache-2.0 HF metadata.

The in-tree MAGI-1 I2V route is wired with the official example image/config assets, local sand-ai/MAGI-1 checkpoint, 4.5B distill profile, sampling controls, fps, and seed.

Running in WorldFoundry

Launch worldfoundry-eval run magi-1 --pipeline.task-profile interactive-video to sample the clip. The command builder above switches --pipeline.task-profile among recorded tasks and still emits worldfoundry-eval run.

It runs in the dedicated environment magi-official-cu124 (Python 3.10.12, CUDA 12.4).

Recorded weight repositories include sand-ai/MAGI-1 and sand-ai/MAGI-1. Prepare the environment and check local assets with bash scripts/inference/prepare_model_infer.sh magi-1. Add --download to fetch missing public checkpoints.

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