Replace Music Section · V4.5

Two numbers on this node decide whether it works: infill_start_s and infill_end_s. The window has to be at least ten seconds wide, and start must be before end. That's it, that's the whole trap - and it's the reason a lot of first attempts at "just fix this one line" fail, because one line is not ten seconds.

What it is

KIE_Next_V4_5_5ef3a450 replaces a time window inside a song you already generated, using Suno's V4.5 provider version through KIE.ai. The pack generates one node per KIE Market operation and provider version, so V4.5 is its own class - the version is pinned in the node, not chosen in a dropdown.

V4.5's own selling point in KIE's notes is "smarter prompts, faster generation, and supports durations of up to 8 minutes", and it's listed as deprecated now, like the rest of the V4 family. In practice that means: use it to repair or extend an existing V4.5-era track, and use the V6 node for new work.

How it works

Infill. You give it the source track's identity, a time window, the new content for that window, and the full lyric sheet as it should read afterwards. KIE re-renders the window in the context of the surrounding audio and returns a complete track - not a stem, not a patch file. The rest of the song is preserved because the model was conditioned on it.

The transport is the pack's standard one: assemble the request, submit, poll until timeout_seconds (1200 by default), download the first result into a ComfyUI AUDIO, and put credits spent and remaining on the node face. Nothing runs on your hardware; Suno is closed weights, which is exactly the job the API keeps winning even with ACE-Step doing good work locally.

Source identification is task_id (the parent generation task) plus audio_id (the specific track inside it). Both are technically optional; both are what tell the provider which song you mean, so in practice wire them. Your music-generation node's task_id socket connects directly, and the audio ID comes out of that task's raw_json.

The fields that matter

The window: infill_start_s and infill_end_s, in seconds, two decimals, start < end, gap ≥ 10. A chorus is usually 15–30 seconds, so a chorus replacement is comfortable; a single word is impossible.

full_lyrics (required) is the complete lyric sheet after the edit - replaced lines plus untouched ones. lyrics (optional) is only the replacement segment. If you're going to get one field wrong on this node, this is the pair to slow down on; sending a full sheet that only contains the new part makes the model think the song is that part.

Required prompt, tags, title: a short description, a style tag string, and a name. negative_tags is the exclusion list and defaults to Rock, so if the replaced section is inexplicably guitar-free you know where to look. variety (0–4, default 1) governs how far the new section may stray; style_weight, weirdness_constraint and audio_weight are 0–1 dials at 0.65 by default, and moving them all at once is how you lose track of what helped.

persona_id/persona_model won't do anything here - KIE documents personas for V5 and newer.

Outputs

audio is a full track: Save Audio, or into KIE • Assemble Video + Music to score a clip. Its task_id is a new task, so you can chain another replacement window. Plus url, all_urls_json, raw_json, credits_consumed, credits_left. Because it isn't an output node, hook the audio to something or the node won't execute at all.

Installing it

ComfyUI Manager → search "KIE.ai Nodes Next", or:

comfy node install kie-ai-nodes-next

Manual:

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

Restart, then Settings → KIE.ai Nodes Next → Connection and paste the key - validated before it overwrites a working one, masked afterwards, and never written to workflow JSON. Environment-variable keys aren't supported. Dependencies are light (requests, Pillow, numpy, PyYAML) and there are no weights to download.

Where people get burned

Window too small. Ten seconds minimum, and the pack's local preflight catches it before upload - a free rejection, which is the good kind.

Full lyric sheet, wrong song. Covered above.

Expecting surgical precision on a four-minute track. V4.5 is the version with the 8-minute ceiling so length isn't your problem here, but the model still rewrites the seam. Listen to the join; if it lurches, rerun with variety lowered rather than stacking a second replacement window on top.

Thinking you can undo it. Each run is billed, and the original is only recoverable if you kept the task_id - which is what KIE • Reuse Completed Image/Video/Audio is for.