Civitai Ace Step Audio

You type a genre description and some lyrics, and the node returns a finished track. Civitai Ace Step Audio runs ACE-Step - the open-weights music model that's the community's local-Suno answer - on Civitai's cloud, with all the knobs the model actually has. No ACE-Step install, no 4GB VRAM requirement, no Sana autoencoder to babysit. Just Buzz.

What it is

A recipe node under Civitai/Audio that submits an aceStepAudio job to the orchestration API. It defaults to ACE-Step 1.5 turbo 2B, which is the version the KB notes dropped the VRAM floor under 4GB - and here you don't even need that. It returns the audio as a native AUDIO blob and, optionally, a VIDEO blob.

Inputs that matter

Required: music_description (the genre/style writeup - the tooltip's example is "Neo-Soul: A warm, organic neo-soul track..."), lyrics (structured with [Verse]/[Chorus]/[Bridge] markers), and seed for reproducible runs.

The useful optional dials:

• duration (default 60s, up to 190), bpm (40–200, default 120), time_signature (2/3/4/6), key (34 choices, mirroring ComfyUI's own TextEncodeAceStepAudio1.5 keyscale), language (23 languages).
• instrumental_weight (default 0.85) vs vocal_weight (default 0.9) - how the model splits its effort. Worth knowing the KB's honest finding here: ACE-Step's instrumentals are the strong suit, vocals and lyrics are the weak suit.
• steps (default 8) and cfg (default 1.0) - and the tooltips carry the important caveat: turbo variants converge at 8 steps / CFG-off at 1.0, but non-turbo base/sft variants expect the full 50-step schedule and CFG around 4. If you swap in a non-turbo diffusion model override and leave steps at 8, you'll get garbage.
• diffusion_model, clip_model, vae_model, language_model - optional AIR overrides if you want something other than the defaults (turbo 2B UNet, qwen_4b_ace15 CLIP, ace_1.5_vae, qwen_0.6b_ace15 language model). Leave them alone unless you know why you're changing them.
• cover_json - supply cover-image config and the output becomes a WebM video with that image as the visual (that's what the blob_video output is for).
• loras - the fun one, because ACE-Step is trainable with LoRAs (artist-style LoRAs, just like checkpoints).

Outputs: blob_audio (AUDIO), blob_video (VIDEO, only populated when cover_json is set), workflow_id, raw_json.

Install & gotchas

Standard pack install - Manager (search Civitai Comfy Nodes), comfy node registry-install civitai-comfy-nodes, or git clone. Auth via env var / OAuth / Civitai Auth node.

The traps, in order:

1. Steps/CFG must match the model variant. This is the biggest one. Turbo defaults (8 steps, CFG 1) are wrong for a non-turbo base model - the tooltips spell out the 50-step / CFG-4 schedule those need.
2. Vocals are the weak suit. Set your expectations: the community's verdict on ACE-Step is "instrumentals genuinely compelling, lyrics genuinely garbage." If you need vocals that carry a song, the local model has the same flaw - it's the model, not the node.
3. It's metered. Music jobs on the cloud cost Buzz, and long durations (up to 190s) cost more. The workflow_id/raw_json outputs exist so you can see the actual cost per run.

One honest note: ACE-Step already runs locally on a 4GB card, so the cloud version's appeal is narrower than for the big image models - it's for when you're on a machine without a card, want the XL/larger variants without the download, or want it wired into a bigger cloud workflow. If you have a half-decent GPU and generate music often, run it local and keep the Buzz.