license: apache-2.0
Magi-1: Autoregressive Video Generation Are Scalable World Models
This repository contains the code for the Magi-1 model, pre-trained weights and inference code. You can find more information on our project page.
1. Introduction
We present magi, a world model that generates videos by autoregressively predicting a sequence of video chunks, defined as fixed-length segments of consecutive frames. Trained to denoise per-chunk noise that increases monotonically over time, magi enables causal temporal modeling and naturally supports streaming generation. It achieves strong performance on image-to-video (I2V) tasks conditioned on text instructions, providing high temporal consistency and scalability, which are made possible by several algorithmic innovations and a dedicated infrastructure stack. Magi further supports controllable generation via chunk-wise prompting, enabling smooth scene transitions, long-horizon synthesis, and fine-grained text-driven control. We believe magi offers a promising direction for unifying high-fidelity video generation with flexible instruction control and real-time deployment.
2. Model and Checkpoints
We provide the pre-trained weights for Magi-1, including the 24B and 4.5B models, as well as the corresponding distill and distill+quant models. The model weight links are shown in the table.
| Model | Link | Recommend Machine |
|---|---|---|
| Magi-1-24B | Magi-1-24B | H100/H800 * 8 |
| Magi-1-24B-distill | Magi-1-24B-distill | H100/H800 * 8 |
| Magi-1-24B-distill+fp8_quant | Magi-1-24B-distill+quant | H100/H800 * 4 or RTX 4090 * 8 |
| Magi-1-4.5B | Magi-1-4.5B (Comming Soon) | RTX 4090 * 1 |
| Magi-1-4.5B-distill | Magi-1-4.5B-distill (Comming Soon) | RTX 4090 * 1 |
| Magi-1-4.5B-distill+fp8_quant | Magi-1-4.5B-distill+fp8_quant (Comming Soon) | RTX 4090 * 1 |
3. How to run
3.1 Environment preparation
We provide two ways to run Magi-1, with the Docker environment being the recommended option.
Run with docker environment (Recommend)
docker pull magi/magi:latest
docker run -it --gpus all --privileged --shm-size=32g --name magi --net=host --ipc=host --ulimit memlock=-1 --ulimit stack=6710886 sandai/magi:latest /bin/bash
Run with source code
# Create a new environment
conda create -n magi python==3.10.12
# Install pytorch
conda install pytorch==2.4.0 torchvision==0.19.0 torchaudio==2.4.0 pytorch-cuda=12.4 -c pytorch -c nvidia
# Install other dependencies
pip install -r requirements.txt
# Install magi-attention, new install method
pip install --no-cache-dir "https://python-artifacts.oss-cn-shanghai.aliyuncs.com/flash_attn_3-3.0.0b2-cp310-cp310-linux_x86_64.whl" --no-deps
3.2 Inference command
# Run 24B Magi-1 model
bash example/24B/run.sh
# Run 4.5B Magi-1 model
bash example/4.5B/run.sh
3.3 Useful configs
| Config | Help |
|---|---|
| seed | Random seed used for video generation |
| video_size_h | Height of the video |
| video_size_w | Width of the video |
| num_frames | Controls the duration of generated video |
| fps | Frames per second, 4 video frames correspond to 1 latent_frame |
| cfg_number | Base model uses cfg_number==2, distill and quant model uses cfg_number=1 |
| load | Directory containing a model checkpoint. |
| t5_pretrained | Path to load pretrained T5 model |
| vae_pretrained | Path to load pretrained VAE model |
4. Acknowledgements
5. Contact
Please feel free to cite our paper if you find our code or model useful in your research.
@article{magi1,
title={Magi-1: Autoregressive Video Generation Are Scalable World Models},
author={Magi-1},
journal={arXiv preprint arXiv:2504.06165},
year={2025}
(TODO: add correct citation)
}
If you have any questions, please feel free to raise an issue.