Suno v6 Wild · Text to Music

Suno v6 Wild is the same deal as the standard v6 node - a style brief in, two finished MP3 songs with vocals out, flat $0.072 per run - with one deliberate difference: the takes are textured and less predictable. Same price, different personality. If the standard model keeps handing you a song that's technically right and completely unmemorable, this is the variant to swap in.

What "wild" buys you

Not chaos, and not a different model with different inputs. The schema is identical to Suno v6 and v6 Mini: same prompt, lyrics, generate_lyrics, instrumental, same dials. The API's own framing is "experimental songs… two textured, less predictable MP3 takes." So you're not unlocking a new mode, you're choosing a model that is more willing to take a left turn - interesting timbres, less radio-shaped arrangements.

In practice that makes it the node you reach for when you're generating material rather than a finished release. Scoring a weird short film, hunting for a texture, trying to get a song that doesn't sound like the other nine hundred songs Suno made today. Reach for standard v6 when you need the hook to land on the first listen.

How the two rules work

The mode logic is shared across the Suno family, and it's where nearly all confusion lives:

• generate_lyrics - the model writes the words from your prompt. The tooltip is blunt about scope: it "only applies when lyrics is empty and instrumental is off."
• lyrics - your words, sung in order. [Verse], [Chorus], [Bridge] and friends shape the structure and are not sung. If you want a specific hook, write it; don't describe it in the prompt and hope.
• instrumental - no vocals at all. Flip it and the prompt becomes pure style.

Then the conditional fields: title is metadata and only applies with lyrics or an instrumental (otherwise the model names it), and the three optional reference sockets - up to five mood images, one video, one audio clip - "only when the model writes the lyrics." Wire a reference in and then type your own lyrics, and you've just unhooked the reference.

Outputs

Two takes come back. The pack folds them into one batched AUDIO on the audio output and Save Audio writes each as its own MP3 file; the second output is task id, the string you quote to support. variety (0 to 4, with 1 as the model's normal) is how far apart the two takes sit - on a "wild" model, raising it is how you get genuinely different directions rather than two shades of the same idea.

style_weight, weirdness, audio_weight and variety all default to -1, which this pack uses to mean "leave the field out, the model decides." Don't read -1 as a low setting.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/spicyapi-ai/comfyui-spicyapi

Restart, then Settings > SpicyAPI and paste a key from spicyapi.ai/console/keys. Headless box? SPICY_API_KEY in the environment is the better route. Restart once after adding the key so the node menu comes from your account's catalogue.

No dependencies to install - the pack declares dependencies = [] and leans on aiohttp, torch, Pillow and PyAV, all of which already ship with ComfyUI. Python 3.10+, V3 node API, tested at ComfyUI 0.39. It'll also install through ComfyUI Manager by searching the pack title.

Where people get burned

The usual API-node realities apply, and they're worth stating once. Every run is billed; nodes ship with control-after-generate set to fixed so a re-queue reuses the cached result instead of silently charging twice. Cancelling only ends your wait - the job finishes on SpicyAPI's side and is charged if it succeeds, with the result sitting in your console for up to 14 days. Put a number in Max cost per run and nothing above it gets submitted.

For this pack specifically, the two errors you're most likely to meet are No SpicyAPI key is set (key missing or not picked up - restart) and 40004, "no deployment serves this combination," which means one of the options you selected isn't served for that model; the message names the field. Anything ending in request_id ... is for support, not for you to debug.

One last thing: Wild is not a secret NSFW switch, and it isn't a licensed-music-workaround either. It's a model variant for texture. If what you actually wanted was a specific voice, that's persona_id on the same node, not a different one.