ACE-Step / acestep-v15-xl-turbo-diffusers

ACE-Step v1.5 XL Turbo (diffusers)

Diffusers-format checkpoint of ACE-Step v1.5 XL Turbo — a guidance-distilled 3.5B-parameter flow-matching DiT for text-to-music generation.

This is the diffusers port of the original ACE-Step release, intended for use with the AceStepPipeline being upstreamed to huggingface/diffusers . The weights are algebraically identical to the upstream release (converted via scripts/convert_ace_step_to_diffusers.py ); the repo layout and scheduler wrapping follow the diffusers convention so the pipeline can be loaded in one call.

Parity vs. the reference ACE-Step implementation on jieyue (bf16 + flash-attn, seed 42, 160 s turbo text2music): waveform Pearson ≈ 1.0, spectral Pearson ≈ 1.0.

Usage

Requires the in-review diffusers PR ( pip install "git+https://github.com/ChuxiJ/diffusers.git@add-ace-step-pipeline" ).

import torch
from diffusers import AceStepPipeline

pipe = AceStepPipeline.from_pretrained(
    "ACE-Step/acestep-v15-xl-turbo-diffusers",
    torch_dtype=torch.bfloat16,
)
pipe.to("cuda")

output = pipe(
    prompt="An upbeat synthwave track with driving drums and a catchy lead",
    lyrics="[Verse]\nNeon lights are calling me\n[Chorus]\nRide the wave tonight",
    audio_duration=30.0,
    generator=torch.Generator(device="cuda").manual_seed(42),
)
audio = output.audios[0]  # (channels, samples) @ 48 kHz

Turbo variants are guidance-distilled and should be run without classifier-free guidance ( guidance_scale=1.0 , the pipeline default for turbo checkpoints) at 8 denoising steps.

Repository layout

├── model_index.json
├── transformer/        # AceStepTransformer1DModel (DiT, 3.5B params, bf16)
├── condition_encoder/  # AceStepConditionEncoder (with baked-in silence_latent)
├── vae/                # AutoencoderOobleck (48 kHz stereo)
├── text_encoder/       # Qwen3-Embedding-0.6B
├── tokenizer/          # Qwen3 tokenizer
├── scheduler/          # FlowMatchEulerDiscreteScheduler config
└── silence_latent.pt   # Raw reference (kept for debugging; not needed at runtime)

License

• ACE-Step weights: MIT (same as upstream )
• text_encoder/ (Qwen3-Embedding-0.6B): Apache 2.0 — redistributed per Qwen's license

Citation

If you use ACE-Step, please cite the original release:

@misc{acestep2025,
  title  = {ACE-Step: A Foundation Model for Music Generation},
  author = {{ACE-Step Team}},
  year   = {2025},
  url    = {https://huggingface.co/ACE-Step}
}