Generate Music · V4

If your ComfyUI workflow needs a song from a text prompt, this is the node - Suno's generate endpoint, wired up as a graph node instead of a curl command, pinned to the V4 model version.

Context for why you're here rather than running a local model: instrumental music is the job local has mostly won (ACE-Step, on embarrassingly small hardware), and vocals plus lyrics are the job where it hasn't. The community's complaint about ACE-Step's lyrics is the same one people make about Suno's, just louder. So this node is the "I need the good version of the singing" choice, and it's metered accordingly.

The two-mode design you have to get right

custom_mode is the hinge of this whole node. On (the default), you're in the detailed-control mode: your prompt is sung verbatim as lyrics, so style and title carry the actual direction, and the fine-tuning weights become live. Off, it's simplified mode: prompt becomes a loose idea that Suno writes its own lyrics from, and KIE wants at least one of an image, video, audio or lyric reference alongside it.

Getting this backwards is the single most common way to waste a generation. Custom mode plus a descriptive sentence is how you end up with a track that sings your description of a piano.

instrumental (default on) is the other half. In custom mode, on means only style and title matter; off means style, title and prompt are all required, with prompt as the exact lyrics.

References - the interesting part

This node takes real media references: image_urls, three numbered audio_1…audio_3 inputs and three video_1…video_3 inputs. They're native ComfyUI sockets, so you can wire a generated image straight in and have it uploaded for you.

There's a documented catch the pack handles for you: KIE only accepts image references when custom mode is off. Older versions of this node would submit the combination anyway and you'd get a provider rejection after paying for the upload. The current pack detects connected images and submits with custom_mode=false, dropping the custom-only controls (duration, style, title, persona and the weights) that the provider would otherwise reject one by one. You'll see it say so in the console. So if you've wired an image in, expect simplified mode whether or not you ticked the box.

The rest of the inputs

prompt (multiline, up to 3000 characters in custom mode on V4), style (the genre, capped at 200 characters on V4 - it's 1000 on V4.5 and later), title (80 characters, required in custom mode, and it ends up in filenames and players, so name it like a human), negative_tags for styles to exclude, and lyrics as a separate field that takes priority over prompt for lyrics. vocal_gender is m or f and only nudges probability. style_weight, weirdness_constraint and audio_weight all run 0–1 and are custom-mode-only. variety (0–4, default 1) is the stylistic spread control. persona_id with persona_model lets you apply a saved style or a cloned voice - that's where the voice-cloning nodes feed in, with persona_model set to voice_persona.

duration shows up too, but KIE documents it as only valid on V5.5 and later; on V4 it does nothing useful, and the versions that support it are the newer ones, not this node.

Outputs, install, and one warning

Outputs are audio (native, wire it to Save Audio), url, all_urls_json, task_id, raw_json, credits_consumed and credits_left. On success the node prints a credits receipt on itself, which is a nice touch when you're iterating.

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

ComfyUI Manager can install it - search "KIE.ai Nodes Next" - then set the key once in Settings → KIE.ai Nodes Next → Connection.

The warning: V4 is marked discontinued in KIE's own documentation, and the node family includes current versions. If you're not chasing a specific V4 sound, run the V6 variant of this node instead. Same wiring, better output.