YuE2 in ComfyUI: workflow, nodes and model files

YuE2 is supported natively in ComfyUI. The ComfyUI documentation requires v0.35.0 or later and recommends v0.35.1 or later on AMD GPUs. Both workflows ship in the Template Library as “YuE2: Text to Music” and “YuE2: Music Cover”; if you cannot find them, your install is older than the required version.

Both files are published in the Comfy-Org/YuE2 repository on Hugging Face. The ComfyUI guide does not state a VRAM figure; the upstream reference runtime specifies a 24 GB NVIDIA GPU, which is covered in the local install guide.

1. Write the Style

   Describe genre, vocal type, language, tempo and instruments. Keep musical instructions here rather than in the lyrics.

2. Enter the Lyrics

   Use section tags such as [Verse], [Chorus], [Bridge] and [Outro], and put only the words that should be sung in this field.

3. Choose a planning Mode

   full writes a melody-and-chord plan before the audio; melody writes the melody only and leaves the accompaniment freer.

4. Set Max Duration

   The value is a ceiling in seconds, not a target. The Text to Music template starts at 120 seconds; the node accepts up to 900.

5. Queue the workflow

   YuE2 writes the plan, generates the song from it, and stops when it writes the end of the song.

YuE2 can write a symbolic plan in ABC notation before it generates audio. That plan is the part you can read and edit, and it is what separates the ComfyUI workflow from a one-shot generator.

The YuE2 Generate Music node also has an advanced cfg_scale option. The templates keep the default of 1.0, which disables guidance and matches the planning workflow.

Max Duration sets a budget of 25 semantic tokens per requested second. Sampling stops as soon as the model writes its end token, so a song whose structure finishes early can be well under the limit.

• A long prompt shortens the song instead of failing: when style, lyrics and the ABC plan use more context than the run can spare, ComfyUI lowers the budget and logs “YuE2 music budget reduced to …”.
• A warning that the semantic stage “reached its token budget before the end token” means the audio can cut off. Raise Max Duration.
• If the same run also logged a budget reduction, the prompt is already using the available context. Shorten the style, lyrics or ABC plan instead.
• When nothing is left, the run stops and asks you to shorten the style, lyrics or ABC.

The cover template does not feed the reference audio to YuE2 directly. SheetSage2 transcribes the melody into an ABC score, and YuE2 generates a new performance and arrangement from that score with your Style and Lyrics. The node default for this workflow is 360 seconds.

Because the link to the original is the melody, covers work best when the new lyrics stay close to the original phrasing and syllable count. Only load recordings you have the right to process.

Not for the official templates. The nodes ship with ComfyUI itself, so a missing node usually means an outdated install or a node that failed to import at startup. ComfyUI Cloud receives the nodes after a stable release.

Community node packs for YuE2 also exist on GitHub. They are separate projects with their own dependencies and model layouts, so read each repository before installing one next to the native nodes.

Use ComfyUI when you want the score, the planning modes and full control of the pipeline. Use the studio when you want to hear whether a set of lyrics and a style brief work before committing a GPU to it.