Generate Music · V4.5PLUS

V4.5PLUS is the top of the V4.5 family - the same Suno generation endpoint as its two siblings, on the version KIE describes as "richer sound, new ways to create, max 8 min". If you're going to spend API credits on a song rather than a loop, this is the node in the legacy row worth picking.

The context that matters: song generation with clean vocals is still the job local models haven't closed. The open answer, ACE-Step, is genuinely compelling on instrumentals and noticeably weaker once someone has to sing. So the API route is a defensible choice here, not just laziness.

What the node does

It builds a Suno request and ships it to KIE.ai under the transport model ai-music-api/generate, with V4_5PLUS as the model version inside the payload. Then it blocks the queue, polling the record-info endpoint every few seconds until the job succeeds or timeout_seconds (default 1200 - twenty minutes) runs out. The returned MP3 is downloaded and decoded locally into a standard ComfyUI AUDIO, using ComfyUI's own torchaudio/PyAV. Nothing runs on your GPU and no weights are involved.

The switches to understand first

custom_mode (default true) changes what your text means: in custom mode prompt is taken as exact lyrics; turn it off and the same words become a description, with Suno writing its own lyrics.

instrumental (default true) removes the vocal. In custom mode with it on, style and title are all the endpoint needs; turn it off and your lyrics get sung.

Non-custom mode is also the only mode where the reference sockets do anything: an IMAGE, up to three AUDIO inputs (audio_1–audio_3), up to three VIDEO inputs (video_1–video_3). The pack handles the upload for you - connect a track and it gets hosted and passed as a URL.

Then the dials, which only apply in custom mode: style_weight, weirdness_constraint, audio_weight (all 0–1, all mild), and variety. persona_id and persona_model are the interesting pair - a persona is a saved style/voice fingerprint you apply to new generations, and you create one with the pack's Generate Persona node.

Outputs

audio is the track - into SaveAudio, or into a video pipeline if you're scoring a render. all_urls_json keeps every URL the response returned; Suno usually produces multiple takes and only the first becomes audio, so this is where the alternates live. url, task_id and raw_json handle linking and forensics - and raw_json is also where you find a track's id, which is what the Extend Music nodes need. credits_consumed and credits_left come back as floats and are printed on the node after every run.

Installing

ComfyUI Manager → search ComfyUI-KIE-Nodes-Next. Or:

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

Restart ComfyUI, then paste your key in Settings → KIE.ai Nodes Next → Connection → Paste / replace KIE.ai API key. The pack validates it against KIE before replacing a working key, keeps it in Comfy's user config instead of the workflow, and shows a masked fingerprint with a connected status. Dependencies: requests, Pillow, numpy, PyYAML - no checkpoints and no CUDA.

The traps

Feeding an image silently rewrites your mode. Whenever image_urls is connected, the pack forces custom_mode=false and strips the custom-only fields (style, title, duration, vocal_gender, style_weight, persona_id, instrumental, variety…) before submitting, with a console message explaining why. It's a guard against a provider-side rejection, but it means your carefully set style disappears the moment you plug in a reference image.

Several widgets are no-ops on this version. duration and persona_model are documented as V5.5/V6-family features. The tooltips say so; nobody reads them.

persona_id defaults to persona_123, a placeholder. Empty optionals are dropped before submission, so clearing it is safe and probably correct.

V4.5PLUS is legacy. KIE tags it (Discontinued) along with V4, V4.5, V4.5ALL, V5 and V5.5; V6, V6_MINI and V6_WILD are the live versions. This node is the right shape on a model version that's on its way out.