README.md

---
license: apache-2.0
language:
- en
pipeline_tag: image-to-video
library_name: MAGI-1
---

![magi-logo](figures/logo_black.png)


-----

<p align="center" style="line-height: 1;">
  <a href="https://static.magi.world/static/files/MAGI_1.pdf" target="_blank" style="margin: 2px;">
    <img alt="paper" src="https://img.shields.io/badge/Paper-arXiv-B31B1B?logo=arxiv" style="display: inline-block; vertical-align: middle;">
  </a>
  <a href="https://sand.ai" target="_blank" style="margin: 2px;">
    <img alt="blog" src="https://img.shields.io/badge/Sand%20AI-Homepage-333333.svg?logo=data:image/svg%2bxml;base64,PHN2ZyB3aWR0aD0iODAwIiBoZWlnaHQ9IjgwMCIgdmlld0JveD0iMCAwIDgwMCA4MDAiIGZpbGw9Im5vbmUiIHhtbG5zPSJodHRwOi8vd3d3LnczLm9yZy8yMDAwL3N2ZyI+CjxwYXRoIGZpbGwtcnVsZT0iZXZlbm9kZCIgY2xpcC1ydWxlPSJldmVub2RkIiBkPSJNMjI3IDIyNS4wODVDMjI3IDIwMi4zMDMgMjI3IDE5MC45MTIgMjMxLjQzNyAxODIuMjExQzIzNS4zMzkgMTc0LjU1NyAyNDEuNTY2IDE2OC4zMzQgMjQ5LjIyNiAxNjQuNDM0QzI1Ny45MzMgMTYwIDI2OS4zMzIgMTYwIDI5Mi4xMjkgMTYwSDUwNy44NzFDNTA5LjI5NSAxNjAgNTEwLjY3NiAxNjAgNTEyLjAxNCAxNjAuMDAxQzUzMi4wODIgMTYwLjAxNyA1NDIuNjExIDE2MC4yNzcgNTUwLjc3NCAxNjQuNDM0QzU1OC40MzQgMTY4LjMzNCA1NjQuNjYxIDE3NC41NTcgNTY4LjU2MyAxODIuMjExQzU3MyAxOTAuOTEyIDU3MyAyMDIuMzAzIDU3MyAyMjUuMDg1VjI1Ni41NThDNTczIDI5MS4zMTkgNTczIDMwOC43IDU2NS4wMzUgMzIzLjI3OUM1NTguNzU2IDMzNC43NzIgNTQzLjU2NSAzNDYuMTEgNTIzLjA3OCAzNTkuNjA1QzUxNC42NzQgMzY1LjE0MSA1MTAuNDcyIDM2Ny45MDkgNTA1LjYzOSAzNjcuOTM2QzUwMC44MDYgMzY3Ljk2NCA0OTYuNTAzIDM2NS4yIDQ4Ny44OTYgMzU5LjY3MUw0ODcuODk2IDM1OS42N0w0NjYuNDY5IDM0NS45MDVDNDU2Ljg3NSAzMzkuNzQyIDQ1Mi4wNzggMzM2LjY2IDQ1Mi4wNzggMzMyLjIxOEM0NTIuMDc4IDMyNy43NzcgNDU2Ljg3NSAzMjQuNjk1IDQ2Ni40NjkgMzE4LjUzMUw1MjYuNzgyIDI3OS43ODVDNTM1LjI5MSAyNzQuMzE5IDU0MC40MzUgMjY0LjkwMyA1NDAuNDM1IDI1NC43OTRDNTQwLjQzNSAyMzguMzg2IDUyNy4xMjUgMjI1LjA4NSA1MTAuNzA1IDIyNS4wODVIMjg5LjI5NUMyNzIuODc1IDIyNS4wODUgMjU5LjU2NSAyMzguMzg2IDI1OS41NjUgMjU0Ljc5NEMyNTkuNTY1IDI2NC45MDMgMjY0LjcwOSAyNzQuMzE5IDI3My4yMTggMjc5Ljc4NUw1MTMuMTggNDMzLjk0MUM1NDIuNDQxIDQ1Mi43MzggNTU3LjA3MSA0NjIuMTM3IDU2NS4wMzUgNDc2LjcxNkM1NzMgNDkxLjI5NCA1NzMgNTA4LjY3NSA1NzMgNTQzLjQzNlY1NzQuOTE1QzU3MyA1OTcuNjk3IDU3MyA2MDkuMDg4IDU2OC41NjMgNjE3Ljc4OUM1NjQuNjYxIDYyNS40NDQgNTU4LjQzNCA2MzEuNjY2IDU1MC43NzQgNjM1LjU2NkM1NDIuMDY3IDY0MCA1MzAuNjY4IDY0MCA1MDcuODcxIDY0MEgyOTIuMTI5QzI2OS4zMzIgNjQwIDI1Ny45MzMgNjQwIDI0OS4yMjYgNjM1LjU2NkMyNDEuNTY2IDYzMS42NjYgMjM1LjMzOSA2MjUuNDQ0IDIzMS40MzcgNjE3Ljc4OUMyMjcgNjA5LjA4OCAyMjcgNTk3LjY5NyAyMjcgNTc0LjkxNVY1NDMuNDM2QzIyNyA1MDguNjc1IDIyNyA0OTEuMjk0IDIzNC45NjUgNDc2LjcxNkMyNDEuMjQ0IDQ2NS4yMjIgMjU2LjQzMyA0NTMuODg2IDI3Ni45MTggNDQwLjM5MkMyODUuMzIyIDQzNC44NTYgMjg5LjUyNSA0MzIuMDg4IDI5NC4zNTcgNDMyLjA2QzI5OS4xOSA0MzIuMDMyIDMwMy40OTQgNDM0Ljc5NyAzMTIuMSA0NDAuMzI2TDMzMy41MjcgNDU0LjA5MUMzNDMuMTIyIDQ2MC4yNTQgMzQ3LjkxOSA0NjMuMzM2IDM0Ny45MTkgNDY3Ljc3OEMzNDcuOTE5IDQ3Mi4yMiAzNDMuMTIyIDQ3NS4zMDEgMzMzLjUyOCA0ODEuNDY1TDMzMy41MjcgNDgxLjQ2NUwyNzMuMjIgNTIwLjIwOEMyNjQuNzA5IDUyNS42NzUgMjU5LjU2NSA1MzUuMDkxIDI1OS41NjUgNTQ1LjIwMkMyNTkuNTY1IDU2MS42MTIgMjcyLjg3NyA1NzQuOTE1IDI4OS4yOTkgNTc0LjkxNUg1MTAuNzAxQzUyNy4xMjMgNTc0LjkxNSA1NDAuNDM1IDU2MS42MTIgNTQwLjQzNSA1NDUuMjAyQzU0MC40MzUgNTM1LjA5MSA1MzUuMjkxIDUyNS42NzUgNTI2Ljc4IDUyMC4yMDhMMjg2LjgyIDM2Ni4wNTNDMjU3LjU2IDM0Ny4yNTYgMjQyLjkyOSAzMzcuODU3IDIzNC45NjUgMzIzLjI3OUMyMjcgMzA4LjcgMjI3IDI5MS4zMTkgMjI3IDI1Ni41NThWMjI1LjA4NVoiIGZpbGw9IiNGRkZGRkYiLz4KPC9zdmc+Cg==" style="display: inline-block; vertical-align: middle;">
  </a>
  <a href="https://magi.sand.ai" target="_blank" style="margin: 2px;">
    <img alt="product" src="https://img.shields.io/badge/Magi-Product-logo.svg?logo=data:image/svg%2bxml;base64,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&color=DCBE7E" style="display: inline-block; vertical-align: middle;">
  </a>
  <a href="https://huggingface.co/sand-ai" target="_blank" style="margin: 2px;">
    <img alt="Hugging Face" src="https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Sand AI-ffc107?color=ffc107&logoColor=white" style="display: inline-block; vertical-align: middle;">
  </a>
  <a href="https://x.com/SandAI_HQ" target="_blank" style="margin: 2px;">
    <img alt="Twitter Follow" src="https://img.shields.io/badge/Twitter-Sand%20AI-white?logo=x&logoColor=white" style="display: inline-block; vertical-align: middle;">
  </a>
  <a href="https://discord.gg/hgaZ86D7Wv" target="_blank" style="margin: 2px;">
    <img alt="Discord" src="https://img.shields.io/badge/Discord-Sand%20AI-7289da?logo=discord&logoColor=white&color=7289da" style="display: inline-block; vertical-align: middle;">
  </a>
  <a href="https://github.com/SandAI-org/Magi/LICENSE" target="_blank" style="margin: 2px;">
    <img alt="license" src="https://img.shields.io/badge/License-Apache2.0-green?logo=Apache" style="display: inline-block; vertical-align: middle;">
  </a>
</p>

# MAGI-1: Autoregressive Video Generation at Scale

This repository contains the code for the MAGI-1 model, pre-trained weights and inference code. You can find more information on our [technical report](https://static.magi.world/static/files/MAGI_1.pdf) or directly create magic with MAGI-1 [here](http://sand.ai) . 🚀✨


## 🔥🔥🔥 Latest News

- Apr 21, 2025: MAGI-1 is here 🎉. We've released the model weights and inference code — check it out!


## 1. About

We present MAGI-1, 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-1 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-1 further supports controllable generation via chunk-wise prompting, enabling smooth scene transitions, long-horizon synthesis, and fine-grained text-driven control. We believe MAGI-1 offers a promising direction for unifying high-fidelity video generation with flexible instruction control and real-time deployment.


## 2. Model Summary

### Transformer-based VAE

- Variational autoencoder (VAE) with transformer-based architecture, 8x spatial and 4x temporal compression.
- Fastest average decoding time and highly competitive reconstruction quality

### Auto-Regressive Denoising Algorithm

MAGI-1 is an autoregressive denoising video generation model generating videos chunk-by-chunk instead of as a whole. Each chunk (24 frames) is denoised holistically, and the generation of the next chunk begins as soon as the current one reaches a certain level of denoising. This pipeline design enables concurrent processing of up to four chunks for efficient video generation.

![auto-regressive denosing algorithm](figures/algorithm.png)

### Diffusion Model Architecture

MAGI-1 is built upon the Diffusion Transformer, incorporating several key innovations to enhance training efficiency and stability at scale. These advancements include Block-Causal Attention, Parallel Attention Block, QK-Norm and GQA, Sandwich Normalization in FFN, SwiGLU, and Softcap Modulation. For more details, please refer to the [technical report.](https://static.magi.world/static/files/MAGI_1.pdf)
<div align="center">
<img src="figures/dit_architecture.png" alt="diffusion model architecture" width="500" />
</div>

### Distillation Algorithm

We adopt a shortcut distillation approach that trains a single velocity-based model to support variable inference budgets. By enforcing a self-consistency constraint—equating one large step with two smaller steps—the model learns to approximate flow-matching trajectories across multiple step sizes. During training, step sizes are cyclically sampled from {64, 32, 16, 8}, and classifier-free guidance distillation is incorporated to preserve conditional alignment. This enables efficient inference with minimal loss in fidelity.


## 3. Model Zoo

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               |
| ----------------------------- | ------------------------------------------------------------ | ------------------------------- |
| T5 | [T5](https://huggingface.co/sand-ai/MAGI-1/tree/main/ckpt/t5) | - |
| MAGI-1-VAE  | [MAGI-1-VAE](https://huggingface.co/sand-ai/MAGI-1/tree/main/ckpt/vae) | - |
| MAGI-1-24B                    | [MAGI-1-24B](https://huggingface.co/sand-ai/MAGI-1/tree/main/ckpt/magi/24B_base)       | H100/H800 \* 8                  |
| MAGI-1-24B-distill            | [MAGI-1-24B-distill](https://huggingface.co/sand-ai/MAGI-1/tree/main/ckpt/magi/24B_distill) | H100/H800 \* 8                  |
| MAGI-1-24B-distill+fp8_quant  | [MAGI-1-24B-distill+quant](https://huggingface.co/sand-ai/MAGI-1/tree/main/ckpt/magi/24B_distill_quant) | H100/H800 \* 4 or RTX 4090 \* 8 |
| MAGI-1-4.5B                   | MAGI-1-4.5B      | RTX 4090 \* 1                   |

## 4. Evaluation

### In-house Human Evaluation

MAGI-1 achieves state-of-the-art performance among open-source models (surpassing Wan-2.1 and significantly outperforming Hailuo and HunyuanVideo), particularly excelling in instruction following and motion quality, positioning it as a strong potential competitor to closed-source commercial models such as Kling.

![inhouse human evaluation](figures/inhouse_human_evaluation.png)

### Physical Evaluation

Thanks to the natural advantages of autoregressive architecture, Magi achieves far superior precision in predicting physical behavior through video continuation—significantly outperforming all existing models.

| Model          | Phys. IQ Score ↑ | Spatial IoU ↑ | Spatio Temporal ↑ | Weighted Spatial IoU ↑ | MSE ↓  |
|----------------|------------------|---------------|-------------------|-------------------------|--------|
| **V2V Models** |                  |               |                   |                         |        |
| **Magi (V2V)** | **56.02**        | **0.367**     | **0.270**         | **0.304**               | **0.005** |
| VideoPoet (V2V)| 29.50            | 0.204         | 0.164             | 0.137                   | 0.010  |
| **I2V Models** |                  |               |                   |                         |        |
| **Magi (I2V)** | **30.23**        | **0.203**     | **0.151**         | **0.154**               | **0.012** |
| Kling1.6 (I2V) | 23.64            | 0.197         | 0.086             | 0.144                   | 0.025  |
| VideoPoet (I2V)| 20.30            | 0.141         | 0.126             | 0.087                   | 0.012  |
| Gen 3 (I2V)    | 22.80            | 0.201         | 0.115             | 0.116                   | 0.015  |
| Wan2.1 (I2V)   | 20.89            | 0.153         | 0.100             | 0.112                   | 0.023  |
| Sora (I2V)     | 10.00            | 0.138         | 0.047             | 0.063                   | 0.030  |
| **GroundTruth**| **100.0**        | **0.678**     | **0.535**         | **0.577**               | **0.002** |


## 5. How to run

### Environment Preparation

We provide two ways to run MAGI-1, with the Docker environment being the recommended option.

**Run with Docker Environment (Recommend)**

```bash
docker pull sandai/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**

```bash
# 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 ffmpeg
conda install -c conda-forge ffmpeg=4.4

# Install MagiAttention, for more information, please refer to https://github.com/SandAI-org/MagiAttention#
git clone [email protected]:SandAI-org/MagiAttention.git
cd MagiAttention
git submodule update --init --recursive
pip install --no-build-isolation .
```

### Inference Command

To run the `MagiPipeline`, you can control the input and output by modifying the parameters in the `example/24B/run.sh` or `example/4.5B/run.sh` script. Below is an explanation of the key parameters:

#### Parameter Descriptions

- `--config_file`: Specifies the path to the configuration file, which contains model configuration parameters, e.g., `example/24B/24B_config.json`.
- `--mode`: Specifies the mode of operation. Available options are:
  - `t2v`: Text to Video
  - `i2v`: Image to Video
  - `v2v`: Video to Video
- `--prompt`: The text prompt used for video generation, e.g., `"Good Boy"`.
- `--image_path`: Path to the image file, used only in `i2v` mode.
- `--prefix_video_path`: Path to the prefix video file, used only in `v2v` mode.
- `--output_path`: Path where the generated video file will be saved.

#### Bash Script

```bash
#!/bin/bash
# Run 24B MAGI-1 model
bash example/24B/run.sh

# Run 4.5B MAGI-1 model
bash example/4.5B/run.sh
```

#### Customizing Parameters

You can modify the parameters in `run.sh` as needed. For example:

- To use the Image to Video mode (`i2v`), set `--mode` to `i2v` and provide `--image_path`:
  ```bash
  --mode i2v \
  --image_path example/assets/image.jpeg \
  ```

- To use the Video to Video mode (`v2v`), set `--mode` to `v2v` and provide `--prefix_video_path`:
  ```bash
  --mode v2v \
  --prefix_video_path example/assets/prefix_video.mp4 \
  ```

By adjusting these parameters, you can flexibly control the input and output to meet different requirements.

### Some Useful Configs (for config.json)

| 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                            |


## 6. License

This project is licensed under the Apache License 2.0 - see the [LICENSE](LICENSE) file for details.

## 7. Citation

If you find our code or model useful in your research, please cite:

```bibtex
@misc{magi1,
      title={MAGI-1: Autoregressive Video Generation at Scale},
      author={Sand-AI},
      year={2025},
      url={https://static.magi.world/static/files/MAGI_1.pdf},
}
```

## 8. Contact

If you have any questions, please feel free to raise an issue or contact us at [[email protected]]([email protected]) .