Upload And Cover Audio · V6

Hand it a track, tell it a style, get a cover back. This is Suno's cover endpoint on V6 - the current tier - and it's the node to use for any new cover work now that KIE's docs list V5 and V5.5 as discontinued.

Why go to an API at all for this? Because covering a song is mostly a vocal job, and vocals are where the local stack thins out. ACE-Step, the open answer to Suno, is competitive on instrumental music and consistently weaker once lyrics and singing are involved. For a cover, the singing is the entire point, so you pay per call and send the source audio off your machine. For one or two covers, fine. For a weekly series, check what a local pipeline gets you first.

How it works

The pack generates one node class per documented model version from KIE's official API documentation at startup, so the tier is baked into the class - V6 here, no dropdown. The node builds the documented request body, submits it to KIE's task endpoint (envelope model ai-music-api/upload-and-cover-audio, inner model V6), polls KIE's record-info endpoint every couple of seconds with backoff until the task completes, then downloads the resulting audio and returns a ComfyUI AUDIO. Failed tasks raise with the task id and the provider's message instead of returning nothing.

The inputs that matter

Three required:

• upload_url - the source track, as a URL KIE can reach. The docs ask that source audio stay under 8 minutes and say the source alone is enough to produce a result.
• prompt - and here's the detail people miss: it's used as lyrics only when lyrics is empty. Fill in lyrics and the prompt stops being lyrics entirely. When instrumental is true, don't pass it at all.
• instrumental - defaults to true, meaning no vocals. That's the opposite of what most people expect from a cover node, so flip it when you want a sung result.

Then: style (up to 1000 characters on this tier), title (80 characters), negative_tags, vocal_gender (m/f; docs concede it only shifts the odds), style_weight, weirdness_constraint, audio_weight, lyrics (5000-character cap, takes priority over prompt), variety (0–4), duration (10–360 seconds, valid on this tier), persona_id and persona_model, plus timeout_seconds (1200) and callback_url.

A small honesty note on persona_model: its tooltip on this node says persona models are only available for "models version 5 and 5.5", which reads like documentation lag rather than a rule - persona IDs come from KIE's persona endpoints and the docs scope them across V5, V5.5 and V6. If personas matter to your workflow, A/B one call before building around it.

Outputs: audio, url, all_urls_json (the response includes both the source and generated URLs - the full set lives here), task_id, raw_json, credits_consumed, credits_left.

Feeding it a track

upload_url is a string widget, so an AUDIO cable won't attach. Add KIE • Upload Audio from KIE Next/Utility/Media, wire your audio into it, then right-click upload_url on this node, choose Convert widget to input, and connect the returned URL. From then on the graph is self-contained: local generate → upload → cover → downstream.

Installing

ComfyUI Manager, search "KIE Nodes Next"; the README still frames registry publication as prepared rather than live, so if nothing shows up there, install by hand:

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 at Settings → KIE.ai Nodes Next → Connection. It's validated before overwriting a working key, only a masked fingerprint reaches the UI, and the app itself never writes the key into your workflow JSON. KIE_API_KEY beats the saved key for headless installs. Dependencies: requests, Pillow, numpy, PyYAML.

Where people get burned

• Instrumental defaults on. Fill in lyrics, get karaoke - that's the default, not a bug.
• Prompt-vs-lyrics precedence. lyrics wins. If your words aren't being sung, look there.
• Local file paths as upload_url - KIE has to fetch it. Upload first.
• No key configured produces an explicit error naming the settings path.
• Mixed-release installs are refused at startup. If version and generated-node build disagree, delete the folder and reinstall.
• Cost of iteration. Every retry is a metered call; credits_consumed is printed on the node so you can see it.
• Rights are your problem. Covering a song you don't own is a licensing question the moment you publish it, and the node has no opinion.

Sequencing that works: nail the cover first, listen to it on its own, and only then feed its audio output into a video or mixing chain. Debugging the song and the picture at the same time just makes both look broken.