Replace Music Section

The one thing local music generation still can't do

ACE-Step is the real story in local music: it generates minutes of audio in seconds, runs on a 4 GB card, is permissively licensed, and treats LoRAs like the image side does. If you want to make music on your own machine, that's the answer, full stop. Its weak suit - and Suno's, for the record - is vocals and lyrics.

What's genuinely hard either way is surgical editing. You have a song. Nineteen seconds in, the second verse is flat and you want it replaced with different lyrics in the same key, same singer, same production. That's not generation, it's infilling into someone else's take, and it's the job this node exists for.

It wraps KIE's Suno replace-section endpoint. You give it the original track's identity, a time window in seconds, and the replacement lyrics, and it returns audio. No retraining, no stem-splitting-and-reassembling in a DAW.

The inputs, in the order they'll trip you up

The origin. task_id and audio_id are optional in the schema, which is misleading - they're how the endpoint knows which song to operate on. Pull them off the task_id and raw_json outputs of your music-generation node. If you're guessing, you're wasting credits.

The window. infill_start_s and infill_end_s are required floats, rounded to two decimals, and the docs are explicit that the window must be at least 10 seconds long. infill_start_s must be less than infill_end_s. Ten seconds is the floor for a reason - you're replacing a musical phrase, not a syllable.

The lyrics. prompt is required and is, per the tooltip, "the lyrics after replacement." full_lyrics is also required and is the complete song text with your new section pasted into the structure - verse markers included. lyrics (optional) is just the replaced segment. Three text fields for one job sounds redundant until you get it wrong once: the endpoint rebuilds the whole song with your section swapped, so it needs the whole song.

The style controls. tags (required, style tags), title (required), then a long optional tail: negative_tags, variety (0–4, default 1 - the tooltip explains 0 as "fully off, strictly the same style" through to experimental), vocal_gender (m/f, and the docs admit this "can only increase the probability"), style_weight, weirdness_constraint, audio_weight, persona_id, and persona_model (style_persona or voice_persona). Plus timeout_seconds and a callback_url to leave alone.

Outputs: audio, url, all_urls_json, task_id, raw_json, credits_consumed, credits_left.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/felipederosilva/ComfyUI-KIE-Nodes-Next
pip install -r ComfyUI-KIE-Nodes-Next/requirements.txt

Restart, then set the key once under Settings → KIE.ai Nodes Next → Connection. The saved key is validated before it replaces a working one, and it never enters workflow JSON.

Practical advice

Do the maths before you run. A 10-second minimum window means you cannot fix a two-second vocal flub with this - for that, you're in a DAW. Work at the song-section scale: a verse, a chorus, a bridge. And keep the original task_id from generation, because both this node and Recovery Audio need it. Cheap habit, saves a re-render.

Suno via a reseller is cloud generation with the usual terms: your lyrics and any reference audio go to someone else's server, and the provider's own moderation applies. For hobby tracks that's fine. For a client deliverable, read the licensing story first - Suno's model has been through enough versions that "which model name did I actually get" is a real question, and the raw_json output is where you'll find it.