MAGI-2-preview
MAGI-2-preview: Scaling Video Generation Models Efficiently
What is it?
What it is
MAGI-2 Preview is an 114B-parameter unified audio-video generation model that activates 6B parameters per token, enabling efficient video generation through a two-stage process (denoising and refinement).
Why it exists
To explore an efficient path for scaling video generation models by optimizing parameter usage and co-designing architecture, systems, and data pipelines.
Who should use it
AI researchers working on video generationMachine learning engineers developing MoE architecturesDevelopers building high-fidelity audio-visual generation pipelinesUsers with high-end NVIDIA Hopper GPU infrastructure
Who should avoid it
Users without NVIDIA Hopper GPUs (8x required for standard inference)Users with limited storage (requires ~307 GB for weights)Users with low VRAM (models require significant offloading/multi-GPU)Users looking for lightweight or mobile-friendly models
How it works
A quick walkthrough in plain English
How MAGI-2-preview works
Step 1 of 3
You interact with it
Open MAGI-2-preview, send a request, or connect it to your stack.
Features
Advantages
- High efficiency with 6B parameters per token despite 114B total
- Unified audio-video generation with synchronized output
- Prompt enhancement improves coherence and quality of generated content
- Open-source under Apache 2.0 license
- Docker and Hugging Face integration simplifies deployment
- Two-stage process enables high-resolution 1080p output
- Supports both text and image prompts for flexible use cases
- Turbo VAE decoder reduces computational overhead during decoding
Disadvantages
- Requires 8 NVIDIA Hopper GPUs for full functionality
- Large model size (114B parameters) demands significant resources
- Currently limited to 10-second video clips
- Prompt enhancement requires API key or custom LLM setup
- Dependency on ffmpeg for audio muxing may limit accessibility
- Large checkpoint size (307 GB total) poses storage challenges
- Preview stage and refiner require separate processing steps
- Hardware-specific (Hopper GPUs) may restrict broader adoption
Installation
docker
docker pull sandai/magi-2-preview:latest\ndocker run --gpus all -it -v /path/to/ckpt:/workspace/ckpt sandai/magi-2-preview:latest
native
pip install -r requirements.txt
FAQ
What hardware and software are required to run MAGI‑2 Preview inference?
MAGI‑2 Preview needs NVIDIA Hopper GPUs (8 of them are recommended for full 1080p runs). It requires Python 3.12, a recent CUDA toolkit, and `ffmpeg` on the PATH for audio muxing. The Docker image `sandai/magi-2-preview` bundles all dependencies, including those that need a compiler.
How do I set up MAGI‑2 Preview using Docker versus building from source?
With Docker, pull the image `sandai/magi-2-preview:latest` and run it with GPU access, mounting your checkpoint directory: ``` docker pull sandai/magi-2-preview:latest docker run --gpus all -it -v /path/to/ckpt:/workspace/ckpt sandai/magi-2-preview:latest ``` If you need to change a dependency version or cannot reach the registry, build locally: ``` docker build -t magi-2-preview:local . ``` From source, install the Python requirements: ``` pip install -r requirements.txt ``` and ensure the `MagiAttention` and `MagiCompiler` repos are available at the pinned revisions.
Where are the model weights and how do I download them?
The weights are not bundled with the code. Download them into a `ckpt/` directory at the repository root: ``` pip install huggingface_hub hf download sand-ai/MAGI-2-preview --local-dir ckpt ``` The directory structure matches the config expectations, so no renaming is needed. You can point to a different location by setting `MAGI2_CKPT_ROOT`.
How can I run a single video generation from the command line?
Use `torchrun` with the entry script and provide the prompt, output directory, and optional image: ``` torchrun --nproc_per_node=8 inference/pipeline/entry.py \ --prompt "a red fox in snow" \ --output output/ ``` Add `--prompt-file`, `--image`, `--seed`, `--config`, `--output-width`, `--output-height`, `--num-inference-steps`, `--refiner-num-inference-steps`, or `--deterministic` as needed. The script writes the MP4 to `output/`.
What is prompt enhancement and how do I enable it?
Prompt enhancement rewrites a short user prompt into a structured, long caption using an LLM, improving model performance. It uses templates in `inference/prompt_enhancement/prompts/`. To enable it, set an OpenAI‑compatible `API_KEY` in `enhancer.py` or subclass `LLMClient`. If `API_KEY` is empty, the raw prompt is used. Example: ``` export API_KEY=sk-... python inference/prompt_enhancement/enhancer.py ```
Featured in Videos
YouTube tutorials and walkthroughs for MAGI-2-preview
Alternatives
Similar projects ranked by category, topics, and text overlap.
