Artistic font generation:Text directly generates music, Meta's new open source model MusicGen explodes!-Font Tutorial免费ppt模版下载-道格办公

On June 13, Meta (the parent company of Facebook, Instagram, etc.) announced the open source of a new language model, MusicGen. Text can be directly generated music.

Except In addition to using text to generate music, MusicGen also supports users to upload sample music to enhance the accuracy of music generation. For example, an upbeat electronic dance track with syncopated drums, airy pads, and intense musical crests. Then upload a similar song "I Can't Stop" and click Generate.

「AIGC Open Community "I tried MusicGen for a while, it is easy to use and powerful, but consumes very little resources. The generated music basically conforms to the text prompt, the sound quality is transparent, the audio is stable, and the sound peak shakes strongly, etc. . In short, the treble is sweet, the mid-range is accurate, and the bass is stable. It is suitable for making background music such as rock, dance, classical, pop, and nostalgic.

Have to generate emotion With the power of traditional AI, ordinary people can use ChatGPT+MusicGen+Gen-2 to make a miniature movie or music MV with only text. The current effect may be slightly worse, but after 5 or 10 years, AI will really subvert the entertainment creation mode by then.

Open source address: https://github.com/facebookresearch/audiocraft

Paper address: https://arxiv.org/abs/2306.05284

Free online test address: https://huggingface.co/spaces/facebook/MusicGen

MusicGen brief introduction

MusicGen is a Text-to-music language models. MusicGen uses a single-stage autoregressive Transformer (ChatGPT also borrows from this model) model. The model is trained with a 32kHz EnCodec tokenizer, 4 codebooks sampled at 50 Hz, and can generate 4 samples at a time.

MusicGen It has three parameters of 300M, 1.5B and 3.3B. If used locally, 16G GPU memory can run.

MusicGen and MusicLM, Riffusion, Musai, and similar products use different techniques that do not rely on self-supervised semantic representation. Instead, conditional on textual description or melodic features, allowing for finer control over musical output conforming to textually prompted criteria.

MusicGen Model trained on 20,000 hours of licensed music data, including entire tracks and individual instrument samples . Specifically, MusicGen relies on an internal dataset of 10,000 hours of high-quality music tracks, as well as music data from ShutterStock and Pond5.

How to use MusicGen

Currently, Meta is in https://huggingface.co/spaces/facebook/MusicGenFree test available The address is very easy to use and can be used directly without registration. (Currently there are a lot of testers, there may be connection error prompts, just refresh a few times)

1) Login test Platform, enter a description in the text field, e.g. 80s electro-rock style with melodic synths, catchy beats, and punchy bass.

2) If you want Enhance the accuracy of music output, you can click to upload similar music. You can also click [Generate] directly without uploading music.

3) Click【 Generate] to generate music.

Comparison between MusicGen and similar products

MusicGen and MusicLM, Riffusion and other similar products have been compared to reflect their powerful functions. For example, strong melodies, tropical percussion and upbeat pop dance music are very suitable for listening on the beach. The following are the music generated by three similar products.

generated from above In terms of music content, MusicGen's music is more transparent, clear, and stable, and its rhythm meets the user's description standard. More music comparison address:https://ai.honu.io/papers/musicgen/

Meta indicates that currently MusicGen is still in the stage of iteration and optimization. In the future, it will open up model training codes and more music functions, thereby subverting the creative process of the music industry.