Zoey - 音乐生成

ComfyUI's audio story is thin, and music generation is the corner where the ecosystem just points at ACE-Step - good at instrumentals, weak at vocals. YuE2 goes the other way: it's a lyrics-forward music model, and ZoeyYue2Studio (「Zoey - 音乐生成」) is one node that takes a one-line description and a duration slider and hands you a finished song.

The whole YuE2 graph is normally eight wired nodes. This collapses it into a single execute() - loading, scoring, sampling, decoding and saving all happen inside. There are no required connections. You type a sentence, drag the slider, hit Run.

It's backend-only: no frontend JS, which is why the style selectors are plain server-side dropdowns.

The one input you actually set

desc is a one-line description of the song. duration is 30–300 seconds, and its tooltip is the most useful sentence on the node: it's an upper bound, not a promise. YuE2 turns it into a token budget at 25 tokens/second, silently shrinks that budget if your style string or lyrics are long, and may finish early on its own. Budget roughly 7 seconds per lyric line - past that, the tail gets truncated. The real length only comes back from the model, so the node prints each take's actual duration in 结果说明.

mode toggles between 文生音乐 (text-to-music) and 翻唱 (cover: transcribe a reference song's melody, then re-sing it with new style and lyrics). Cover needs reference audio - either the optional ref_audio AUDIO input (which wins) or ref_audio_file. No reference, no cover: it fails early rather than wasting a run.

takes (1–4) gives you multiple versions, seeds spaced base + i × 7919 - one take isn't enough to judge vocals.

Then the style selectors: language, vocal, genre, bpm, instrument_1–instrument_3 and production, composed server-side into one style string, same formula as YuE Studio. Two details matter: the language must be explicitly selected or you'll get Chinese lyrics sung with an English accent, and the BPM needs its unit. style overrides the whole thing if you fill it in. lyrics takes [Verse]/[Chorus] tags; leave it blank with ai_lyrics on and a local llama.cpp server writes it from desc. Priority everywhere is manual over automatic: your style and lyrics always win, and the AI only fills what's empty.

ckpt_name picks the YuE2 checkpoint, lora_name an optional adapter (applied to both MODEL and CLIP - style and melody tokens come out of the CLIP half), and format is flac or mp3.

Two outputs. 音频 (AUDIO) is the take; wire it to a save-audio node. 结果说明 (STRING) carries per-take durations plus warnings - send it to Show Text instead of leaving it dangling.

Three optional extras: vc_enabled swaps the lead vocal for the reference audio's timbre, asr_enabled transcribes the finished song to check the words landed, and llm_model picks the GGUF used for lyric writing.

Install

ComfyUI Manager, or:

cd ComfyUI/custom_nodes
git clone https://github.com/liangzoey/comfyui-ZoeyTool.git
cd comfyui-ZoeyTool
pip install -r requirements.txt

Restart. Torch, Pillow, numpy and opencv come from that; the YuE2 model itself is core ComfyUI, supplied separately.

The traps, in order of how often they'll bite you

1. It needs a ComfyUI new enough to have YuE2 in core. run() imports comfy_extras.nodes_yue2, the SheetSage2 encoder from nodes_audio_encoder and comfy_api.latest.UI. It's registered unconditionally, so on an older build it appears in the menu and dies with an ImportError on Run. Update ComfyUI first.

2. Non-Windows users: the helper engines are hardcoded to the author's machine. Absolute Windows paths for llama-server, the CUDA bin directory, the ASR environment and Seed-VC's Python sit at the top of nodes/zoey_yue2_studio.py. On Linux or macOS the llm_model dropdown shows 「（没找到 GGUF 模型）」 and AI lyrics, voice conversion and transcription can't run until you edit those constants. Text-to-music doesn't need them.

3. Name your files so the dropdown can find them. ckpt_name filters for checkpoints with yue2 in the filename, then falls back to listing every checkpoint if nothing matches - so it's easy to pick a Qwen-Image checkpoint and get a nonsense error. Same fallback on the LoRA list, and cover mode wants a SheetSage2 audio encoder in models/audio_encoders.

4. Pin the seed and it silently does nothing. No IS_CHANGED anywhere in this pack. With seed on control_after_generate (the default) each Run regenerates; pin it and ComfyUI serves cached audio, skipping the AI-writing and voice-conversion subprocesses entirely.

5. The CPU fallback has no error message. llama.cpp needs the CUDA bin directory on PATH; without it the server silently falls back to CPU - ten times slower, and it just looks like the node is being slow.

With ai_lyrics or vc_enabled on, GPU work is serialized so two models are never resident at once. Takes land in output/yue/<timestamp>/ with an mp3/ subfolder, lyrics.txt and meta.json.