Sign In

MiniMax Music 3 12GB — caption + lyrics to a song with vocals, int8 DiT, mp3 out

Download

1 variant available

Config Other

The_frizzy1_minimax-music3-12gb_v1.0.0.json

23.95 KB

Verified:

Type
Workflows
Stats

9

Reviews
Published

Nov 21, 2026

Base Model

Other

Hash
AutoV260C68758A3

☕ One coffee = one happy creator: https://buymeacoffee.com/the_frizzy1


MiniMax Music 3 on a 12 GB card. You write a caption and lyrics, it gives you a song with vocals as an mp3. One workflow file, core ComfyUI nodes only. No video on this workflow yet — the one above is how to use my workflows, if you are new to them.

What you need

✔ Custom nodes — None. Core ComfyUI only.
✔ ComfyUI version — 0.42 or newer. On anything older the nodes are not there.
✔ VRAM — 12 GB — 11.5 GB peak, tested, not estimated
✔ System RAM — 32 GB · ComfyUI itself peaked at 4.3 GB
✔ Tested on — RTX 3060 12 GB · 32 GB RAM · ComfyUI 0.43.1 · sage attention
✔ Licence — Workflow MIT · MiniMax Music 3: Apache-2.0

Models to download

Click a file name to download it. Put it in the folder shown, inside ComfyUI/models/.

Required — 3 files, total 11.9 GB

diffusion_models/

File: minimax_music3_dit_int8_convrot.safetensors — 2.50 GB (music)
Download:
https://huggingface.co/Comfy-Org/MiniMax-Music-3/resolve/main/diffusion_models/minimax_music3_dit_int8_convrot.safetensors

Put it here:
ComfyUI/models/diffusion_models/

text_encoders/

File: minimax_music3_text_encoder_pruned_int8_convrot.safetensors — 9.20 GB (music)
Download:
https://huggingface.co/Comfy-Org/MiniMax-Music-3/resolve/main/text_encoders/minimax_music3_text_encoder_pruned_int8_convrot.safetensors

Put it here:
ComfyUI/models/text_encoders/

vae/

File: minimax_music3_dav.safetensors — 0.22 GB (music)
Download:
https://huggingface.co/Comfy-Org/MiniMax-Music-3/resolve/main/vae/minimax_music3_dav.safetensors

Put it here:
ComfyUI/models/vae/

How to fill it in

caption: genre, BPM, key, mood, instruments. lyrics: sections in brackets, [Verse], [Chorus] — leave it empty for an instrumental. max_duration: seconds.

Speed on my RTX 3060

Wall is queue to finished. VRAM is the whole card as nvidia-smi reports it, about 0.7 GB of that is Windows.

  • 60 s of music, first run after starting ComfyUI — 60 s, 2 min 20 s wall, 11.3 GB peak vram, 3.6 GB peak ram (comfyui)

  • 120 s of music — 120 s, 4 min 28 s wall, 11.5 GB peak vram, 4.3 GB peak ram (comfyui)

What to watch out for

  • 11.5 GB peak on a 12 GB card. There is almost nothing left. Close everything else that touches the GPU.

  • Time scales with the length you ask for. 120 s cost roughly double the 60 s run, so a four minute track is a long wait.

  • It is lighter than the official template only because it uses the int8 DiT (2.5 GB) instead of fp16 (4.9 GB). If you compare quality with the official one, that is the difference you are hearing.

  • Only two runs are measured, both from the template's own lo-fi caption. Nothing was tested across genres.

Get the workflow

More videos on this stuff: https://www.youtube.com/playlist?list=UUGKDBMhTGqvlBZE9cF5d50Q