SandAI-org/MAGI-2-preview

MAGI-2-preview: Scaling Video Generation Models Efficiently

What it solves

MAGI-2 Preview is designed to scale video generation efficiently. It addresses the challenge of producing high-resolution (1080p) video with synchronized audio from text or image prompts, while optimizing the computational cost of running a massive model (114B parameters) by using a Mixture-of-Experts (MoE) architecture that only activates a small fraction of its parameters per token.

How it works

The model uses a two-stage generation process: the magi2_preview stage denoises a clip at low resolution, and the magi2_refiner stage upscales the result to 1080p. It supports both text-to-video (T2V) and image-to-video (I2V) workflows. To improve output quality, it includes an optional prompt-enhancement step where an LLM rewrites short user prompts into structured, detailed captions.

Who it’s for

This project is for researchers and developers with high-end hardware (specifically 8 NVIDIA Hopper GPUs) who want to generate high-quality, 10-second video clips with integrated audio.

Highlights

  • Unified Audio-Video Generation: Generates both video and sound simultaneously, muxing them into a single output file.
  • Efficient Scaling: Utilizes a 114B parameter model that only activates 6B parameters per token.
  • High Resolution: Capable of producing 1080p video via a two-stage preview-and-refine pipeline.
  • Flexible Input: Supports both text-to-video and image-to-video generation.

Related

  • Project
    SandAI-org/MAGI-1

    MAGI-1 is an autoregressive video generation model that produces high-fidelity videos chunk-by-chunk to ensure temporal consistency and physical accuracy.

  • Project
    HITsz-TMG/Uni-MoE

    Uni-MoE is a Mixture-of-Experts based omnimodal large model capable of understanding and generating across multiple modalities, including text, images, and speech.

  • Project
    apiframe-ai/seedance-2.0-api

    An API implementation and example set for ByteDance's Seedance 2.0, enabling text, image, and multimodal reference-to-video generation with synchronized audio.

  • Project
    open-gigaai/giga-world-1

    An open-source framework for training and deploying robot world models to evaluate robot policies through controllable video generation.

  • Project
    inclusionAI/Ming

    Ming-flash-omni 2.0 is an open-source omni-MLLM that unifies multimodal perception and generation across text, image, audio, and video in a single MoE-based architecture.