metadata

license: cc-by-nc-nd-4.0
arxiv: 2311.11255
extra_gated_prompt: >-
  Please fill in the following fields, the full name/institution/group/contact
  email/use case are MUST fields, and gender/github/personal homepage are
  OPTIONAL fields (You can simply use a '-' symbol to fill in these optional
  fields). An application form without required information will be declined.
extra_gated_fields:
  Full Name: text
  Gender: text
  Institution: text
  Group: text
  Contact Email: text
  Github: text
  Personal Homepage: text
  Use Case: text
  I agree to use this dataset for non-commercial use ONLY: checkbox
tags:
  - music

MUImage Dataset

This is the MUImage dataset used to facilitate image to music generation, consisting of 9,966 music files with a total playtime of 27.72 hours generated using the MU-LLaMA and BLIP Image Captioning models. This dataset is used to train the M²UGen model.

The MUImageInstructions.json file contains a list with each of the element having the following format:

{
  "input_file": "1OeqDC6vmto.jpg",
  "output_file": "1OeqDC6vmto.mp3",
  "conversation": [
    {
      "from": "human",
      "value": "Generate a music for the image that is ethereal and spacey, with a focus on the synthesizers and bass, and a hint of a pulsing beat.",
      "input_modality": "image",
      "caption": "the image shows a bright star in the center of a galaxy"
    },
    {
      "from": "gpt",
      "value": "Here is a music that is space-themed and could fit well with the image of a bright star in the center of a galaxy.",
      "caption": "The music is a blend of electronic and acoustic instruments, with a focus on the synthesizers and bass.",
      "output_modality": "audio"
    }
  ]
}