WorldFoundry
English中文

MAGI-2 Preview

MAGI-2 Preview from sand-ai: text-to-video and image-to-video; Planned.

Edit this pageGitHubMarkdown

MAGI-2 Preview is a video generation model from sand-ai. The catalog records it for text-to-video and image-to-video. It is also cataloged under the aliases magi2, magi2_preview, and magi-2-preview.

No official teaser or method figure is checked into this docs tree.

Code: SandAI-org/MAGI-2-preview.

Build your run

Pipeline
NativeMagi2Pipelineworldfoundry.pipelines.magi2.pipeline_magi2
TaskDefault default
Environmentworldfoundry-unified-cu128
DeviceCUDA 12.8
Runnerworldfoundry.pipeline
Loadingfrom_pretrained
Invocationnative_or_standard_invocation
worldfoundry-eval zoo model-download --model-id magi2-preview --check-local --json

What it is for

magi2-preview is the recorded text/image-to-video line. Use it when you want both prompt and still conditioning on the same catalog card.

Catalog tasks include text-to-video and image-to-video. Catalog aliases include magi2, magi2_preview, and magi-2-preview. Documented modalities: text and image in, video out.

MAGI-2 preview T2V/I2V. Not MAGI-1.

Runtime contract

Required prompt and image; optional video, audio, and actions (seed, task); artifacts magi2-preview.mp4.

Recorded task profiles: text-to-video and image-to-video. WorldFoundry binds NativeMagi2Pipeline.

Family status: Planned.

114B-parameter MoE (256 experts, top-6), ~6B active per token; joint audio+video, 32/44.1kHz muxed.

Ported faithfully from SandAI MAGI-2-preview (Apache-2.0), single-GPU PyTorch (cp/ep/dp collapsed).

Upstream targets 8xHopper; this port collapses cp/ep/dp to single-GPU (sequential residency for the 114B MoE).

Running in WorldFoundry

The catalog launch is worldfoundry-eval run magi2-preview --pipeline.task-profile default. 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).

Recorded weight repositories include sand-ai/MAGI-2-preview and sand-ai/MAGI-2-preview. Prepare the environment and check local assets with bash scripts/inference/prepare_model_infer.sh magi2-preview. Add --download to fetch missing public checkpoints.

Recorded licenses include apache-2.0; catalog VRAM 80 GB; status: Planned.