Mochi 1: The Top Open Source Video Generation Tool You Need to Try

Genmo’s Mochi 1 is a video generation model that creates high quality video with exceptional motion realism from a text prompt.

Setting a new standard for video generation, Mochi is a tool that is optimized for photorealistic styles and generates realistic motion.

Released under the permissive 2.0 license, users can modify the model and make nearly any change to better suit their problem. The model features a 10 billion parameter diffusion model built on the Asymmetric Diffusion Transformer (AsymmDiT) architecture which efficiently processes user prompts alongside compressed video tokens allowing for high quality video generation.

Because Mochi was trained from scratch, there is better motion fidelity, prompt adherence, and visual quality compared to similar tools trained on existing models. In this article, we'll explain what Mochi is, how to run it on serverlessly on Beam, and some of the popular use-cases it can be used for.

Getting Started with Mochi 1

Let’s take a look at how you can use Mochi with Beam. For help getting started with Beam checkout the documentation.

Upload Model Weights

First, you need to upload the model weights to Beam. This is handled by the upload.py script:

Then run the script. This will upload the Mochi weights to Beam's storage volumes, so you can use them for inference.

Setup The Environment

Setup your remote environment and define the dependencies in the mochi_image:

Inference Function

Now, we'll create function to generate video from text prompt in your app.py file:

Deployment and Inference

We can deploy the model to Beam by running this command.

When your app is deployed, you'll get a URL you can use to make inference requests.

Send a post request to your API endpoint, including your prompt like this:

We'll get back a file with the video generated. Not bad!

For a more in depth tutorial, checkout Beam’s Mochi tutorial here.

Evaluations and Performance

As an open-source video generation model, Mochi 1 outperforms other video AI models in internal tests. The model performs best when provided with a very detailed text prompt like the following example:

"The camera follows behind a rugged green Jeep with a black snorkel as it speeds along a narrow dirt trail cutting through a dense jungle. Thick vines hang from towering trees with sprawling canopies, their leaves forming a vibrant green tunnel above the vehicle. Mud splashes up from the Jeep’s tires as it powers through a shallow stream crossing the path. Sunlight filters through gaps in the trees, casting dappled golden light over the scene. The dirt trail twists sharply into the distance, overgrown with wild ferns and tropical plants. The vehicle is seen from the rear, leaning into the curve as it maneuvers through the untamed terrain, emphasizing the adventure of the rugged journey. The surrounding jungle is alive with texture and color, with distant mountains barely visible through the mist and an overcast sky heavy with the promise of rain."

The user can control characters, adjust settings like the camera angle, and create or modify actions in the generated video with their prompt.

Competing with other top restricted-access models, Mochi 1 is a great choice for high-quality video generation, particularly in photorealistic styles.

Currently the model can generate videos up to 5.4 seconds; however, Genmo plans to release model changes to allow for longer video generation.

Applications and Use Cases

Mochi is a versatile tool that can be used for many applications such as:

• Education: AI-generated videos can be used in corporate training, medical simulations, and more to create realistic instructional content.
• Entertainment: An artist with minimal funds could use Mochi 1 to create visuals for their music at a low cost.
• Marketing: A startup could use the model to create short video ads for Instagram and TikTok.

There are many more uses for Mochi 1 and its capabilities beyond traditional video generation.

Developer Resources and Community

Mochi 1 is available on the Hugging Face platform, allowing developers to easily access and integrate the model. Because the model is open-source, users can easily use the model, modify the code with minimal restrictions, and ultimately develop a model that suits their specific needs. Currently Genmo is allowing users to preview version 1 of Mochi using their hosted playground and the company is working to expand their team of developers in order to advance the state of video generation and continue improving Mochi 1. You can find more information about Genmo and Mochi 1 here.

Future Vision and Roadmap

Genmo plans to release Mochi 1 HD later this year, which will support 720p resolution and offer even greater motion fidelity. While Mochi 1 is an advanced text to video tool, Genmo plans to continue developing and improving Mochi. In particular, Genmo wants to create a model/s that support high resolution and allow for longer video generation as well as image-to-video capabilities. Seeing such great results with only their first model, Genmo hopes to continue providing high quality and accessible text-to video tools like Mochi.