Easy SongGeneration - 生成伴奏

Easy SongGeneration - 生成伴奏 ("Generate BGM") is the pure-music mode of this pack: accompaniment, backing track, instrumental - no vocals. The community verdict on open music models has been that instrumentals are their strong suit (ACE-Step's instrumentals get praised precisely where its vocals get laughed at), and Tencent's model has a dedicated mode for it that this node exposes. It's the one you reach for when the goal is a soundtrack, a backing loop, or a song you can hum along to.

How it's different under the hood

The inputs look identical to the mixed node - 歌词, 描述, seed, duration, sampling knobs, auto-style or reference audio - but read the code and the difference shows up in two places. First, your 歌词 input is discarded: the model gets . as the lyric text, because there's no one to sing. Second, your 描述 gets wrapped with [Musicality-very-high], [Pure-Music], before the model sees it. That [Pure-Music] tag is the switch that tells the upstream pipeline to produce instrumental audio.

So yes - the lyrics box on this node is a trap. You can leave it empty or dump anything in there; it does nothing. Spend your effort on 描述 instead. That's where the genre, instruments and vibe live: cinematic, strings, slow build, orchestral is a very different track from lo-fi, warm piano, vinyl crackle.

The inputs that matter

• 描述 (descriptions) - the actual creative control. Defaults to female, pop, energetic, piano, drum kit, which is a strange default for a pure-music node; change it or you'll get instrumentals with "female pop" energy. (Lightly annoying, worth knowing.)
• 种子 / 时长 - -1 random / time-based seed; 0 = the model's max_dur from config (2m30s on base, 4m30s on large/v2).
• 自动参考风格 - genre presets like Electronic, Jazz, Soundtrack that inject a matching reference prompt. Needs the auto-prompt weights; None skips them.
• 参考音频 (prompt_audio) - an optional ComfyUI AUDIO reference that overrides the auto style. Works without the auto-prompt weights, but the pack needs the Demucs files (part of the third_party/ download) to separate your reference into usable parts.

Outputs are 音频 (48kHz AUDIO) and 元数据 (JSON with the seed and generation params).

When to use it vs. the others

If you only need the backing track, this is the cleaner choice - no vocals to strip later, and it's the same GPU cost as any other generation. If you think there's even a chance you'll want the vocals too, the Easy SongGeneration - 分轨生成 node runs one generation and hands back mixed + vocal + BGM together, so you don't pay twice. BGM is the one to reach for when you're sure: YouTube background music, game sound, or a bed under someone else's voiceover.