Empty Ace Step 1.0 Latent Audio

Ace Step is the open music-generation family that made local songwriting feel realistic - text to music, plus reference-audio "timbre" so you can clone a voice or a style into a new track. This node is the empty latent for the 1.0 line: the box the sampler fills with a song.

What it creates

[batch_size, 8, 16, length] where length = seconds × 44100 / 512 / 8. So the 44.1kHz audio is compressed in two stages - 512 samples per unit at 16 frequency bins - into a latent that's surprisingly small: 8 channels, 16 bins, and roughly 10.8 latent steps per second of audio. The node tags it type: "audio" so downstream nodes treat it as sound, not image.

The shape difference from the generic EmptyLatentAudio is the tell: that one is 64 wide with a 2048 temporal downscale, this one is 8x16 with a 4096 downscale. Different audio codecs, different latent shapes. Feed the wrong empty latent to an Ace Step model and the sampler will either error or produce noise that never turns into music.

Where it sits

Text-to-music in ComfyUI goes: prompt → TextEncodeAceStepAudio → this empty latent → sampler → VAEDecodeAudio → audio out. For timbre transfer you additionally feed a reference audio latent so the model borrows the voice. The seconds knob is your runtime budget: Ace Step was designed to be fast on modest hardware - people routinely generate a song in under 20 seconds on mid-range cards - but every second is more latent steps, and a 4-minute track at this ratio is a real job for the sampler.

The knob that matters

seconds (default 120, up to 1000) and batch_size (default 1). That's it. The 1.0 line's relationship to the 1.5 line is worth knowing before you pick: 1.5 uses a different audio codec (64 channels, 1764 downscale) and has its own empty-latent node. They are not interchangeable. If your workflow references an Ace Step checkpoint, match the empty latent to that version - 1.0's node here, 1.5's node over there.

What's not obvious

This family is genuinely recent - Ace Step 1.0 landed in mid-2025 and the ComfyUI integration followed, with 1.5 arriving in early 2026 and driving most of the community's attention since. The model got traction precisely because it runs locally on hardware that video models laugh at, and because reference-audio cloning plus LoRA training made it useful rather than a novelty. If you're coming from image/video work, the mental model flips: the "image" is a spectrogram-like latent, the "prompt" is song structure, and the empty latent is where the song grows. Start with the default 120 seconds, sample something short, and only then push the length.