V2MIDI Dataset

Overview

The V2MIDI dataset pairs 40,000 MIDI files with AI-generated videos, connecting music and visual art in a new way. It's designed to help researchers and artists explore how to synchronize music and visuals using AI. This dataset isn't just a collection of files – it's a tool that could change how we create and experience audio-visual content.

How We Created the Dataset

We built the V2MIDI dataset through several key steps:

Gathering MIDI Data: We started with a large archive of drum and percussion MIDI files, focusing on house music. We picked files based on their rhythm quality and how well they might match with visuals.
Standardizing MIDI Files: We processed each chosen MIDI file to make a 16-second sequence. We focused on five main drum sounds: kick, snare, closed hi-hat, open hi-hat, and pedal hi-hat. This helped keep things consistent across the dataset.
Linking Music to Visuals: We created a system to turn MIDI events into visual changes. For example, a kick drum might make a peak of strength in the visuals, while hi-hats might make things rotate. This is the core of how we sync the music and visuals.
Creating Visual Ideas: We came up with 10,000 text prompts across 100 themes. We used AI to help generate ideas, then went through and refined them by hand. This gave us a wide range of visual styles that fit well with electronic music.
Making the Videos: We used our MIDI-to-visual system and tools such as Parseq, Deforum and Automatic1111 (Stable Diffusion web UI) to create videos for each MIDI file.
Organizing and Checking: Finally, we paired each video with its MIDI file and organized everything neatly. We carefully made sure the visuals matched the music well and looked good.

Why It's Useful

The V2MIDI dataset is special because it precisely matches MIDI events to visual changes. This opens up some exciting possibilities:

See the music: Train AI to create visuals that match music in real-time.
Hear the visuals: Explore whether AI can "guess" the music just by watching the video.
New creative tools: Develop apps that let musicians visualize their music or let artists "hear" their visual creations.
Better live shows: Create live visuals that perfectly sync with the music.

Flexible and Customizable

We've built the V2MIDI creation process to be flexible. Researchers and artists can:

Adjust how MIDI files are processed
Change how music events are mapped to visual effects
Create different styles of visuals
Experiment with video settings like resolution and frame rate
Adapt the process to work on different computer setups

This flexibility means the V2MIDI approach could be extended to other types of music or visual styles.

Training AI Models

One of the most important aspects of the V2MIDI dataset is its potential for training AI models. Researchers can use this dataset to develop models that:

Predict musical features from video content
Create cross-modal representations linking audio and visual domains
Develop more sophisticated audio-visual generation models

The size and quality of the dataset make it particularly valuable for deep learning approaches.

How to Get the Dataset

The dataset is quite big so we've split it into 257 parts of about 1GB each. Here's how to put it back together:

Download all the parts (they're named img2img_part_aa to img2img_part_jw)
Stick them together with this command: cat img2img_part_* > img2img-images_clean.tar
Unpack it: tar -xvf img2img-images_clean.tar

Make sure you have at least 257GB of free space on your computer for this.

What's Next?

We see the V2MIDI dataset as just the beginning. Future work could:

Include more types of music
Work with more complex musical structures
Try generating music from videos (not just videos from music)
Create tools for live performances

Thank You

We couldn't have made this without the people who created the original MIDI archive and the open-source communities behind Stable Diffusion, Deforum, and AUTOMATIC1111.

Get in Touch

If you have questions or want to know more about the V2MIDI dataset, email us at: [email protected]