How to Run Minimax Music 3 Locally in ComfyUI

TL;DR

Run Minimax Music 3 locally by updating ComfyUI, loading its text-to-music template, and placing the required diffusion model, text encoder, and VAE in their designated model folders. The generator works offline for free and supports styles including pop, punk rock, R&B, progressive house, jazz, metal, and country, with model variants designed for different hardware capacities.

Transcript

The best open-source AI music generator is here. It's called Minimax music and not only can this create super clean and great sounding songs, but it's also incredibly tiny. The smallest model is only 2.5 GB in size, so this can fit on both consumer hardware. In this video, we're going to go over some demos and how to install it so you can run it fo... Read More

Key Insights

• Minimax Music 3 is an open-source AI music generator that can run locally through ComfyUI, allowing free and unlimited offline generation after the required files are installed.
• The generator supports a broad musical range, with demonstrations covering pop rock, punk rock, R&B and neo-soul, EDM progressive house, jazz, metal, and country songs.
• The diffusion model is available in multiple sizes, including a 9.8 GB full model, a 4.9 GB FP6 version, and a 2.5 GB int8 Convo version for different hardware capacities.
• The installation requires a diffusion model, a text encoder, and a VAE, with each file placed in its corresponding ComfyUI models subfolder before refreshing the model list.
• The smallest listed int8 text encoder is 9.2 GB, while the workflow's single VAE is 217 MB, so the complete installation requires more storage than the diffusion model alone.
• The Minimax workflow currently supports text-to-music generation only, which means it cannot directly perform music-to-music transformations or create song covers within this workflow.
• The recommended prompt structure has three sections: global metadata for genre and musical character, vocal details for voice qualities, and arrangement instructions for instruments and song sections.
• The demonstrations show that Minimax can generate sung vocals in different languages and reproduce stylistic vocal traits, including the Southern accent heard in the country example.

Questions & Answers

Q: How do you install Minimax Music 3 in ComfyUI?

Update ComfyUI to the latest version, start the application, open Templates from the left sidebar, and search for music. Select the Minimax text-to-music workflow. Download a diffusion model, a text encoder, and the VAE, then place them in the corresponding diffusion_models, text_encoders, and VAE folders inside ComfyUI's models directory.

Q: What files are required to run Minimax Music 3 locally?

The ComfyUI workflow requires three types of files: a Minimax diffusion model, a compatible text encoder, and the provided VAE. The demonstrated options include diffusion models sized at 9.8 GB, 4.9 GB, and 2.5 GB. The smallest listed int8 text encoder is 9.2 GB, and the VAE is 217 MB.

Q: How should a Minimax Music 3 prompt be structured?

A Minimax song description should be divided into three sections. Start with global metadata covering genre, overall feeling, pace, and key. Next, describe the vocals, including gender, pitch, smoothness, coarseness, and timbre. Finish with arrangement details such as instruments and the intended structure of the intro, verses, bridge, and outro.

Q: What music genres can Minimax Music 3 generate?

The demonstrations show Minimax Music 3 generating pop rock, punk rock, R&B and neo-soul, EDM progressive house, jazz, metal, and country. These examples include both instrumental styling and sung lyrics. The country demonstration also shows a Southern vocal accent, while another example demonstrates that the system can sing in a language other than English.

Q: Can Minimax Music 3 run offline for free?

Minimax Music 3 is presented as a free, open-source music generator that can run locally and offline through ComfyUI. Once ComfyUI and the required model files are installed, songs can be generated without a stated usage limit. The practical hardware and storage requirements depend on which diffusion model and text encoder variants are selected.

Q: Which Minimax Music 3 model should you download?

The workflow provides several diffusion model choices for different hardware capacities. The full model is 9.8 GB, the FP6 model is 4.9 GB, and the int8 Convo model is 2.5 GB. The tutorial selects the FP16 model, but users with less available capacity can consider one of the smaller listed variants.

Q: Can Minimax Music 3 create covers or transform existing songs?

The demonstrated ComfyUI workflow currently provides text-to-music generation only. It does not offer music-to-music processing, so it cannot directly transform an existing track or create a cover through that workflow. The description indicates that editing or inpainting songs and making or stacking loops are discussed later, but no detailed procedures appear in the supplied transcript.

Q: How do you make ComfyUI recognize the downloaded Minimax models?

Place every downloaded file in its designated ComfyUI models subfolder: the diffusion model goes in diffusion_models, the text encoder goes in text_encoders, and the VAE goes in VAE. After all files have finished downloading, press R to refresh the model list, then choose each newly installed file from its corresponding dropdown in the workflow.

Summary & Key Takeaways

• Minimax Music 3 is presented as a free, open-source generator that can create complete songs locally through ComfyUI. The demonstrations cover pop rock, punk rock, R&B and neo-soul, progressive house, jazz, metal, and country, including vocals in different languages and a Southern country accent.

• Installation begins by updating ComfyUI and opening the Minimax text-to-music template from the template browser. The workflow requires three components: a diffusion model placed in the diffusion_models folder, a text encoder placed in text_encoders, and a 217 MB VAE placed in the VAE folder.

• Song prompts should describe global metadata, vocal qualities, and arrangement. Global metadata can specify genre, mood, pace, and key, while vocal instructions can define pitch, texture, and timbre. Arrangement instructions can identify instruments and organize the intro, verses, bridge, and outro. The available workflow currently supports text-to-music only.