Zhenzhen Suno Description

Suno's description mode is the laziest way to make a song that actually sounds like what you meant. You write a vibe instead of a spec - "a wistful synthwave track about a road trip that never happened" - and Suno's LLM fills in lyrics, structure, and arrangement from that one sentence. Comfly_suno_description is that feature as a ComfyUI node: description_prompt in, two finished AUDIO takes out, ready to save.

Like the rest of the Comfyui_Comfly pack it's a paid relay call to ai.comfly.chat - no local model, API key required, billed per generation.

What you set

• description_prompt - the whole trick. Vague descriptions produce generic music; specific ones - genre, mood, tempo, instruments, lyrical theme - produce something you'd actually show a friend. This is the one input that determines whether the node feels magical or useless.
• title - your song title (can be empty, but a good one influences Suno's lyrics).
• version - v3.0 through v5 (default v4.5). Each maps to a different Suno backend under the hood (chirp-auk, chirp-bluejay, chirp-crow for the modern ones).
• make_instrumental - flip this for a pure instrumental, no vocals.
• seed - set it to reproduce a take, since Suno otherwise rolls the dice every run.

What comes out

The classic Suno two-fer: audio1 and audio2 are both AUDIO objects (save via a core audio node or VHS), audio_url1/audio_url2 the hosted links, clip_id1/clip_id2 ready to feed into Comfly_suno_cover or Comfly_suno_upload_extend. Uniquely, this node also echoes back the prompt it actually used, plus the tags and title Suno settled on - handy if you want to see how the description mode interpreted your sentence. task_id and response for tracking and errors.

Install

ComfyUI Manager → search "Comfyui_Comfly" → install → restart, or:

cd ComfyUI/custom_nodes
git clone https://github.com/ainewsto/Comfyui_Comfly

API key from ai.comfly.chat into the optional apikey field; it's stored in the pack's Comflyapi.json and reused by the other nodes.

Where people get burned

The failure mode is all upstream of the node: a thin description_prompt gets you a thin song, and you'll re-run it four times chasing a good one (each run costs a little). Write the prompt like a director's note, not a haiku. Also remember both takes are generated together - if the first one slaps and the second is trash, don't re-run the whole node; just use audio1. And because the whole thing rides the relay, a refund-on-failure is the extent of the safety net if the service hiccups mid-generation. Treat it as a jukebox with opinions: describe, generate, keep the good take.