ACE-Step 1.5 lands as an open-source text-to-music foundation model, generating full songs in under 10 seconds on consumer GPUs through ComfyUI.
-
ACE-Step 1.5 is an open-source text-to-music foundation model from ACE Studio and StepFun, available on Hugging Face under Apache-2.0.
-
Pairs a Language Model planner with a Diffusion Transformer decoder for chain-of-thought song structure planning.
-
Generates full songs in under 2s on A100, under 10s on RTX 3090, at 48 kHz stereo quality.
-
Supports vocals in 50+ languages, LoRA fine-tuning, cover generation, repaint, and vocal-to-BGM editing tasks.
-
XL variant uses a 4B-parameter DiT decoder and scores 4.79 on musical coherence, beating most commercial competitors.
-
Natively integrated in ComfyUI via Template Library → Audio with an all-in-one workflow. ACE Studio and StepFun have released ACE-Step 1.5, an Apache-2.0 music generation model now available through an official ComfyUI workflow. The model weights generate 48 kHz stereo songs with optional vocals, giving developers a local alternative to hosted services such as Suno and Udio.
According to the ComfyUI guide, ACE-Step 1.5 can generate a complete song in under two seconds on an Nvidia A100 or under 10 seconds on an RTX 3090. Track length, sampling settings, software versions, and hardware configuration can affect those results.
The official workflow accepts two primary inputs:
- Style description: Genre, instruments, tempo, mood, vocal qualities, and other musical directions.
- Lyrics: Optional text organized with labels such as
[Verse]and[Chorus]. An empty lyrics field produces an instrumental.
One prompt, two-stage synthesis #
ACE-Step 1.5 separates composition planning from audio rendering. A language model first derives metadata, captions, song structure, and other conditioning information from the prompt and supplied lyrics. A Diffusion Transformer, or DiT, then converts that plan into a compressed audio representation.
The stack uses AutoencoderOobleck as its audio variational autoencoder. It maps 48 kHz stereo audio into 64-channel latent data at 25 frames per second, equivalent to one latent time step for every 1,920 audio samples. After diffusion finishes, the autoencoder expands that compact representation into the final waveform.
A fixed Qwen3-Embedding-0.6B text encoder produces 1,024-dimensional embeddings for the prompt. These numerical representations give the planner and diffusion model information about the requested style, structure, and lyrics.
This story is for Pro members #
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.