SongBloom: Coherent Song Generation via Interleaved Autoregressive Sketching and Diffusion Refinement

</div>

We propose SongBloom, a novel framework for full-length song generation that leverages an interleaved paradigm of autoregressive sketching and diffusion-based refinement. SongBloom employs an autoregressive diffusion model that combines the high fidelity of diffusion models with the scalability of language models. Specifically, it gradually extends a musical sketch from short to long and refines the details from coarse to fine-grained. The interleaved generation paradigm effectively integrates prior semantic and acoustic context to guide the generation process. Experimental results demonstrate that SongBloom outperforms existing methods across both subjective and objective metrics and achieves performance comparable to the state-of-the-art commercial music generation platforms.

Models

| Name | Size | Max Length | Prompt type | 🤗 | | -------------------- | ---- | ---------- | ----------- | -------------------------------------------- | | songbloom_full_150s | 2B | 2m30s | 10s wav | link | | songbloom_full_150s_dpo | 2B | 2m30s | 10s wav | link | | songbloom_full_240s$^{[1]}$ | 2B | 4m | 10s wav | link | | ... | | | | |

[1] For the _150s series models, each [intro], [outro], and [inst] corresponds to an expected duration of 1 second; whereas for the _240s series models, each token corresponds to 5 seconds (details in docs/lyric_format).

Updates

Oct 2025: Release songbloom_full_240s; fix bugs in half-precision inference ; Reduce GPU memory consumption during the VAE stage.
Sep 2025: Release the songbloom_full_150s model with DPO post-training
Jun 2025: Release the songbloom_full_150s and inference script

Getting Started

Prepare Environments

conda create -n SongBloom python==3.8.12
conda activate SongBloom

# yum install libsndfile
# pip install torch==2.2.0 torchaudio==2.2.0 --index-url https://download.pytorch.org/whl/cu118 # For different CUDA version
pip install -r requirements.txt

Data Preparation

A .jsonl file, where each line is a json object:

{
	"idx": "The index of each sample", 
	"lyrics": "The lyrics to be generated",
	"prompt_wav": "The path of the style prompt audio",
}

One example can be refered to as: example/test.jsonl

The prompt wav should be a 10-second, 48kHz audio clip.

For details on lyric formatting, see docs/lyric_format.md.

Inference

source set_env.sh

python3 infer.py --input-jsonl example/test.jsonl


# For GPUs with low VRAM like RTX4090, you should set the dtype as bfloat16
python3 infer.py --input-jsonl example/test.jsonl --dtype bfloat16

# SongBloom also supports flash-attn (optional). To enable it, please install flash-attn (v2.6.3 is used during training) manually and set os.environ['DISABLE_FLASH_ATTN'] = "0" in infer.py:8

model-name: Specify model version, see the model cards (eg: songbloom_full_150s/songbloom_full_150s_dpo);
local-dir: Dir where the weights and config files are downloaded;
input-jsonl: input raw data;
output-dir: Dir where the output audio saved;
n-samples: How many audios will be generated for each input term;

Mac Silicon

Set these environment variables before running:

export PYTORCH_ENABLE_MPS_FALLBACK=1
export DISABLE_FLASH_ATTN=1

When loading the model, explicitly pass the MPS device and use float32, not bfloat16:

import torch

device = torch.device('mps')
model = SongBloom_Sampler.build_from_trainer(cfg, strict=False, dtype=torch.float32, device=device)

Citation

@article{yang2025songbloom,
title={SongBloom: Coherent Song Generation via Interleaved Autoregressive Sketching and Diffusion Refinement},
author={Yang, Chenyu and Wang, Shuai and Chen, Hangting and Tan, Wei and Yu, Jianwei and Li, Haizhou},
journal={arXiv preprint arXiv:2506.07634},
year={2025}
}

License

SongBloom (codes and weights) is released under the LICENSE.

SongBloom

Install / Use

README