Replace Music Section · V4.5ALL

The interesting control on this node isn't the one you'd expect. Everyone fiddles with the timestamps first; the number that actually decides whether your replacement belongs in the song is variety, a 0–4 dial that ranges from "same style, exactly" to "unreasonably varied" in the tooltip's own words. Get that right and the ten-second window rule becomes the only mechanical thing left.

What it is

KIE_Next_V4_5ALL_438fc3b1 regenerates a slice of an existing Suno track, pinned to the V4.5ALL provider version, via KIE.ai. One node per version is how the KIE Next pack works - the pack reads KIE's catalog and provider docs and generates a ComfyUI node for each operation/version combination, which is why the class name is a hash and why there are nine nodes in the Replace Music Section folder.

Two honest notes about this specific variant. KIE describes V4_5ALL and V4_5 with identical words - smarter prompts, faster generation, up to 8 minutes of track - so the practical difference between this node and its V4.5 sibling is thin. And the whole V4 family is deprecated in KIE's model notes, superseded by V6, V6_MINI and V6_WILD. Keep this one for reproducing older work.

How it works

Section replacement is infill on the provider side: identify the source track, name a time window, describe what should now be there, and the model re-renders that stretch conditioned on the audio around it. You get a whole track back, not a stem or a diff.

Under the hood it's the pack's standard generated-node pipeline - build the request, submit to KIE, poll until timeout_seconds (1200 default), download the first result URL into a ComfyUI AUDIO, then display the credit receipt. Your task_id comes out as a socket so you can chain runs.

The inputs that matter

Source first: task_id (parent task) and audio_id (the specific track in it). Both optional in the schema, both effectively required - the tooltips describe exactly those roles, and without them the provider doesn't know what it's replacing. Wire the generate node's task_id output, and pull the audio ID out of its raw_json.

Then the window: infill_start_s and infill_end_s, seconds with two decimals, start < end, minimum ten seconds of width. That constraint is not negotiable and it's why "just redo this one phrase" doesn't work here.

Then lyric bookkeeping: full_lyrics (required) is the whole song as it should read after the replacement; lyrics (optional) is just the replaced part. Required prompt, tags and title are description, style tags and name - and note that KIE's prompt tooltip says "the lyrics after replacement", which is a doc slip that has confused a lot of people into pasting verse text there.

Now the knobs worth your attention. variety: 0 is "strictly the same style", 1 balanced, 2 clearly different, 3 bold exploration, 4 explicitly "may differ in style very significantly". For a section that must blend, stay at 0–1. For "give me an alternate bridge", 2–3. style_weight, weirdness_constraint and audio_weight are 0–1, all defaulting to 0.65, and my advice is to move one at a time - they interact and you'll never attribute the result otherwise. negative_tags defaults to Rock, which does have an effect. vocal_gender nudges, doesn't guarantee. persona_id/persona_model do nothing on a V4-family pin.

Outputs

audio → Save Audio / Preview Audio, or KIE • Assemble Video + Music for the video-scoring case. task_id seeds a follow-up replacement. url is remote and temporary, all_urls_json lists everything, raw_json is the provider's own answer, credits_consumed and credits_left are the receipt. And remember it's not an output node - no downstream consumer, no execution.

Installing it

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

ComfyUI Manager search "KIE.ai Nodes Next" and the Registry package kie-ai-nodes-next work too. Restart, then Settings → KIE.ai Nodes Next → Connection to paste your key; it's checked against KIE before replacing a saved one and only a masked fingerprint is ever shown. Env vars are unsupported on purpose. Deps: requests, Pillow, numpy, PyYAML. No weights, no VRAM.

Where people get burned

Sending a full_lyrics sheet that only contains the new section - the model then treats that as the entire song.

Setting variety to 4 on a section that has to sit between two you like, and being surprised when the bridge arrives from a different album.

And the cost model: every replacement is a paid render, good or bad. Right-click → KIE Next · Check inputs & model controls inspects the cached documented limits without uploading or submitting, which is the free way to sanity-check a window before you spend.