Extend Music · V4.5PLUS

A track that stops at 90 seconds when it wanted to be three minutes is the normal outcome of AI music generation, and this node is the fix for it. Feed it the ID of a Suno track you already generated, point it at the spot where the extension should begin, and Suno's V4.5PLUS model keeps going - same style, same mood, new material. The result lands back on your canvas as a regular ComfyUI AUDIO.

The ID is the gate

audio_id is required and it is not a URL - it's the provider's identifier for a track Suno itself made, the id field inside the response. You get it from a Generate Music node's raw_json output. Paste a link to your own MP3 here and it fails, because the extend endpoint has no idea what to do with one; that's what the Upload And Extend Audio node is for.

Practically, this makes Extend Music the second half of a chain: generate → grab the ID you like → extend it. Extending the result again is how people with patience and credits assemble a full-length song out of one good chorus.

How it runs

Remote request to KIE.ai, transport model ai-music-api/extend, V4_5PLUS as the version in the payload. The node then blocks - polling until the job completes or timeout_seconds (default 1200 seconds) expires - and decodes the returned audio locally with ComfyUI's own torchaudio/PyAV.

continue_at (default 60, in seconds) is the dial that decides everything. Set it at the tail of the original and you get a clean continuation; set it early and you're asking the model to re-perform the rest of the track, which sometimes rescues a mediocre middle section and sometimes destroys one.

prompt is required as a widget and serves as the extension's lyrics - unless lyrics is filled, which overrides it. style should mirror the source, because a mismatched style makes the join audible. instrumental defaults to false on this node, so an instrumental will get vocals unless you turn it on - and when you do, clear prompt and vocal_gender, since the docs say neither should be sent in instrumental mode.

Then the usual trim: negative_tags, variety, and the style_weight / weirdness_constraint / audio_weight nudges. duration and persona_model are V5.5/V6-family features per the docs, so they're decorative here.

Outputs

audio is the extended track - the output you actually want, ready for SaveAudio or a video pipeline. all_urls_json retains every URL the provider returned (there's usually more than one take; audio is the first), url is the primary link, task_id is KIE's job handle for the detail/inspection utility nodes, and raw_json carries the full payload - that's where the next audio_id lives if you're chaining extensions. credits_consumed and credits_left come back as floats and are also displayed on the node after the run.

Install

ComfyUI Manager → search ComfyUI-KIE-Nodes-Next. Or the manual route:

cd ComfyUI/custom_nodes
git clone https://github.com/felipederosilva/ComfyUI-KIE-Nodes-Next

Restart, then visit Settings → KIE.ai Nodes Next → Connection and paste your KIE.ai key. It's checked against KIE before being saved, stored in Comfy's user config instead of your workflow JSON, and shown as a masked fingerprint with a connected status. Headless installs can set KIE_API_KEY instead and skip settings entirely. The pack's dependency list is requests, Pillow, numpy, PyYAML - no models, no CUDA.

Where it bites

Extends cost credits and you can't preview first. Auditioning five different continue_at values is five paid generations. Decide the cut first, then run once.

V4.5PLUS is the discontinued rung. KIE's parameter docs group it with V4, V4.5, V4.5ALL, V5 and V5.5 as (Discontinued), while V6, V6_MINI and V6_WILD are current. This node is the right shape, on a version that's on its way out - the V6 sibling in the same folder is the one to build on.

Extension isn't arrangement. If the source is a loop, you'll get more loop. If you want a real second section with new instrumentation, you're better off either writing it yourself in a DAW or generating a fresh track and cross-fading.