Empty Ace Step 1.5 Latent Audio

The Ace Step 1.5 line upgraded the music-generation codec, and this node is the matching empty latent - the box you hand the sampler when the checkpoint is 1.5, not 1.0. If you're choosing between the two Ace Step empty-latent nodes, this is the one for the newer model, and the difference in latent shape tells you exactly what changed under the hood.

What it creates

[batch_size, 64, length] where length = seconds × 48000 / 1920. The 1.5 codec runs at 48kHz, and the formula works out to 25 latent steps per second of audio - 40ms per step - with the node tagging the latent type: "audio" and declaring a temporal downscale ratio of 1764 for the decoder side.

Compare that to the 1.0 node, which is 8x16 wide with a 4096 downscale. 1.5 traded the small multi-dimensional shape for a much wider 64-channel single axis - the same "wider is richer" shift that image models made when they went from 4 latent channels to 16. More channels means less of the audio signal gets thrown away, which is why 1.5 handles reference timbre and cover-style edits better than 1.0 did. It's also why the two nodes are emphatically not interchangeable: the tensor shapes don't match, and the sampler will reject the wrong one.

Where it sits

Same skeleton as every Ace Step workflow: prompt → text encode → this empty latent → sampler → VAEDecodeAudio. The seconds default is 120 (up to 1000), with batch_size alongside. For the 1.5 line, reference-audio conditioning is a first-class feature - you can hand it a timbre reference so a track borrows a voice or style, and the wider latent is a big part of why that works without the artifacts the 1.0 line showed.

What people actually hit

Ace Step 1.5 arrived in early 2026 and is where the community's attention went - the 1.0 line is effectively the older generation. Two practical notes from the field. First, quality-vs-length is real: 1.5 runs fast on modest cards, but a long track is a long latent, and at 25 steps/second a 5-minute song is a serious sampling job. Second, the 1.5 model got a reputation for occasionally dropping or flubbing lines - it's a hit-miss ratio, not a guarantee - so treat the first pass as a draft, not a master. The empty latent itself never misbehaves; it's just a correctly-sized zero tensor. The behavior is all in what you ask the sampler to fill it with.

The one-liner

1.0 latent → 1.0 model. 1.5 latent → 1.5 model. Match them, don't mix them, and you're on solid ground. Ships with ComfyUI core - the Ace Step integration lives in comfy_extras/nodes_ace.py.