Update, then open the template
ComfyUI’s documentation requires version 0.35.0 or later and recommends 0.35.1 or later for AMD GPUs. Update ComfyUI, restart it and search the Template Library for YuE2: Text to Music.
You can also open the official workflow JSON. Native templates do not require a third-party YuE2 custom-node pack. Community studio interfaces are separate installations.
Put the checkpoint in the right folder
The current native template uses yue2_3b_int8_convrot.safetensors from Comfy-Org/YuE2. The SheetSage2 file below is needed for the cover template.
ComfyUI/
└── models/
├── checkpoints/
│ └── yue2_3b_int8_convrot.safetensors
└── audio_encoders/
└── sheetsage2_bf16.safetensors # coversRefresh model selection or restart after placing the files. If the model list is empty, check the directory and filename before downloading another copy.
Make the first song easy to diagnose
- In Style, describe language, genre, vocal character, tempo and instruments.
- In Lyrics, use section labels such as
[Verse]and[Chorus]. - Use full planning for a melody with chords, or melody for freer accompaniment.
- Set a modest Max Duration and queue the workflow.
In this ComfyUI workflow, leaving the abc input empty disables the planning input. The included planning templates wire it for you. Max Duration is a ceiling, not a promise that a song will run for exactly that length.
Fix the common setup problems
Missing nodes: check the installed ComfyUI version and startup import errors. An old native installation is not repaired by randomly adding custom nodes.
Out of memory: close other GPU jobs, start with a shorter song and follow the template’s memory settings. The official Python baseline and this INT8 workflow have different memory behavior.
Words being sung as instructions: move production notes out of Lyrics and into Style. Our prompt builder separates them.
Sources & further reading
- ComfyUI · YuE2 guide ↗
- ComfyUI · Text to Music workflow ↗
- Comfy-Org · YuE2 models ↗
- ComfyUI · Music Cover workflow ↗
Documentation checked against the linked sources. We have not benchmarked generation on our own hardware.