VRGDG YuE2 Generate Song

Type a style, paste lyrics, queue, get a 48 kHz stereo song. No credits, no API, no upload. That's the pitch, and it's real - YuE was the first local music model the community stopped treating as a toy ("Locally? Probably YuE" was the standing answer when ACE-Step was still young), and YuE2 is its current generation.

The catch has followed YuE since launch day: it's a 3B LLM-style music model, and the community's first reaction to the original was the VRAM requirements saying "use 80 GB, tensor parallelism helps." People did get it running on consumer cards with quantization and offloading, but treat a 12 GB card as a diagnostic target, not a destination. The pack's runtime note puts the official target at an NVIDIA BF16 GPU with 24 GB of VRAM, Linux-first, native Windows experimental.

The mechanism, because it isn't one pass

YuE2 generates a song in stages, and knowing the stages explains almost every setting on this node.

First it plans: lyrics and style go through the model to produce an ABC score - ABC being the text notation for melodies, where G2 A2 B2 c2 is real pitches with durations, plus chord symbols if you asked for them. That's the symbolic skeleton of your song, and it's editable text.

Then it generates: semantic tokens from the plan, acoustic latents from those, and a VAE decode into 48 kHz stereo audio. This second half is the expensive one, and it's the only half ode_steps touches.

That split is why this pack has three generation nodes instead of one: you can plan, look at the score, hate it, and replan without paying for synthesis. See VRGDG YuE2 Create Plan and VRGDG YuE2 Render ABC.

Inputs worth setting

yue2_config is the typed bundle from the installer or settings node - you don't type into it, you wire it.

style is a comma-separated description: language, genre, vocal character, instruments, production feel, BPM. The tooltip is explicit that lyrics don't belong here, and it's right - mixing the two is the fastest way to get mush. style_preset gives you 35 prepared prompts (five each for rock, pop, country, rap, hip hop, metal and 90s alternative) that fill the box for you; edit the text and it flips back to Custom / Keep typed style and uses your wording.

lyrics is the actual words, tagged [Verse], [Chorus], [Bridge], [Outro]. Keep sections short.

Then the three dials that matter:

• planning_mode - full plans melody and chords, melody plans without chords, off skips symbolic planning entirely. Use full for original songs, melody for covers where you want to change the harmony, off when you're feeding it something else.
• ode_steps - default 32, which is YuE2's official quality setting. 16 is faster and audibly worse; 48–64 is slower with diminishing returns. It does not make planning smarter or longer, only the final render.
• cfg_scale - leave it at -1 for the model's own default. Pushing to roughly 1.2–1.4 can shove the result harder toward your style prompt, at a quality cost. It's a lasso, not a magic wand.

seed is there for A/B comparison, with the usual caveat from the tooltip: same seed plus different prompts still gives you different audio. filename_prefix names the output folder.

Outputs

audio is the song - a normal ComfyUI AUDIO you wire into a save/preview node, or into a music-video graph downstream. abc_score is the notation of what it sang; feed it to Render ABC later to re-perform the same melody. artifact_directory is where everything landed - a timestamped folder under ComfyUI/output/Yue2 holding audio, ABC, semantic tokens, latents, config and timing. metadata_json is the run summary as a string. And truncated is a boolean you should actually look at: true means the plan ran past the model's length limit and your lyrics got clipped.

Install

Get the pack via Manager (search vrgamedev) or:

cd ComfyUI/custom_nodes
git clone https://github.com/vrgamegirl19/comfyui-vrgamedevgirl.git

Then the real work: add VRGDG YuE2 Installer + Settings, point target_root at a dedicated folder, and click Install Generation + Models. The pack's own requirements don't include YuE2 at all - the model, the 3B weights and the VAE live in an isolated virtualenv the installer builds, because YuE2 pins a different Torch than ComfyUI. Queueing the graph never installs anything. The bundled Workflows/Yue2/VRGDG_YuE2_Starter.json already has the installer's yue2_config wired into this node, so you only edit style and lyrics.

Where people get burned

• Every run reloads the model. In isolated_process mode each run spawns a fresh worker that loads YuE2 from disk, generates, and exits. No cached VRAM, but no free lunch either. That's the price of not sabotaging your ComfyUI environment.
• VRAM contention. If a video model is already sitting on the GPU, this will OOM. unload_comfy_models is on by default for exactly this reason, and offload_ar moves unused autoregressive modules to system RAM if you're still tight.
• Windows attention weirdness. Some Windows Torch wheels advertise Flash Attention without a compiled CUDA kernel, which used to crash YuE2's CUDA-graph capture. The pack patches around it - if you hit attention errors, update the pack and restart instead of chasing your Torch install.
• Long songs get worse, not just slower. More sections means more tokens, longer plans and more truncation. Start with a verse and a chorus.

If you mainly want instrumental music fast on modest hardware, ACE-Step is the cheaper answer. YuE2 is the one that takes lyrics seriously - and charges VRAM for it.