Upload And Extend Audio · V4.5ALL

This is the sibling of the Extend Music node with one crucial difference: it doesn't care who made the track. Point it at an audio file you already have - your own recording, a Suno export, a stem you mixed in a DAW - tell it where to pick up, and let Suno's V4.5ALL model carry on from there.

How it works

Remote call, same shape as everything else in this pack. The node assembles the "upload and extend" request, posts it to KIE.ai, and then sits in the queue polling for a result until timeout_seconds (default 1200, twenty minutes) is exhausted. KIE's transport model is ai-music-api/upload-and-extend-audio, and V4_5ALL is passed as the model version in the payload. The finished audio comes back decoded as a standard ComfyUI AUDIO object - the audio analog of getting an IMAGE back from an API image node.

The inputs you'll actually touch

upload_url is required and it wants a publicly reachable URL, not a file path. In the current build this is a plain string widget, so if you just generated a track three nodes upstream, you'll need to get it online before KIE can read it - the pack ships upload helpers under KIE Next / Utility for that. There's a hard limit worth internalising: keep the source under 8 minutes.

continue_at is the real dial - the second where the extension should begin, default 60. Put it at the tail of the track. Put it in the middle and Suno will effectively re-perform the rest, which can be a feature (a fix for a bad bridge) or a bug (it quietly rewrites the bit you liked).

prompt is required as a widget and is used as lyrics for the extension, but only when lyrics is empty - lyrics takes priority. style should match the source; title and negative_tags are cosmetic and exclusion lists respectively. style_weight, weirdness_constraint, audio_weight and variety behave as elsewhere: nudge-y, optional, and only occasionally decisive.

instrumental is the trap. It defaults to true, and the docs are explicit that when it's on you should not pass prompt, lyrics or vocal_gender at all. It also changes what counts as "done": with it on, the source alone is enough. Clear the prompt box rather than leaving the default sentence sitting there if you flip it.

What comes out

Seven outputs. audio is the decoded extension - so is the output the whole extended song or only the new section? The tooltip describes the result as the extended audio track, and the provider's response also carries a source_audio_url alongside the new audio_url, so in practice you're extending a track, not just producing a tail. Treat all_urls_json as your safety net: it holds every URL the response returned. url, task_id and raw_json are the debugging trio, and credits_consumed / credits_left tell you what the run cost and what's left in the account (also rendered on the node itself after each run).

Installation

ComfyUI Manager → search ComfyUI-KIE-Nodes-Next ("KIE.ai Nodes Next"). Or manually:

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

Restart ComfyUI. Dependencies are requests, Pillow, numpy, PyYAML - this pack downloads zero model weights. Then Settings → KIE.ai Nodes Next → Connection → Paste / replace KIE.ai API key; the key is checked against KIE before it's stored, kept out of workflow JSON, and shown masked. KIE_API_KEY beats it for server installs.

Where it bites

That 8-minute ceiling is a rejection, not a trim. Send a 12-minute file and the task fails rather than being cut down.

Version status. V4.5ALL sits in the (Discontinued) group in KIE's parameter docs (with V4, V4.5, V4.5PLUS, V5, V5.5); V6/V6_MINI/V6_WILD are the current family. The V5 and V4.5PLUS variants of this same node exist in the same folder, and if you're choosing fresh, pick a V6 sibling.

Extending is not genre transfer. If you want the source re-sung or re-styled, that's Upload And Cover Audio - a different node, different endpoint, different result.

Blocking behaviour. The node holds the queue while it waits. Running it inside a long video batch means everything downstream of the audio step stalls for minutes at a time; that's expected, not a hang.