What Is SongBloom
SongBloom is a full-length song generation framework developed by Tencent AI Lab, combining autoregressive sketch drawing with diffusion-based refinement techniques. Through an interleaved generation paradigm, it alternately generates semantic and acoustic context to produce high-quality complete songs. The model requires only a 10-second audio sample and corresponding lyrics as input to generate up to 2 minutes and 30 seconds of dual-channel, 48kHz audio. SongBloom performs excellently in audio quality and lyrics accuracy, approaching the state of the art (SOTA), and has been successfully open-sourced.
SongBloom's Main Features
-
Efficient Song Generation: With only a 10-second audio sample and corresponding lyrics, it can generate a complete song up to 2 minutes and 30 seconds long, supporting dual-channel, 48kHz high-quality audio output.
-
Innovative Generation Paradigm: It adopts an interleaved generation paradigm, combining autoregressive sketch drawing and diffusion-based refinement techniques to alternately generate semantic and acoustic context, optimizing the overall song structure and sound quality.
-
Outstanding Sound Quality and Accuracy: It performs excellently in audio quality and lyrics accuracy, approaching the state of the art (SOTA) and surpassing existing open-source models.
-
Open Source and Ease of Use: The project has been open-sourced, providing detailed usage guides and multiple model versions, supporting operation on low-VRAM devices, and making it easy for users to get started quickly.
-
Broad Application Prospects: It provides a powerful tool for fields such as music creation and audio production, significantly improving creative efficiency and inspiring new ideas in music creation.
SongBloom's Technical Principles
-
Interleaved Generation Paradigm: By alternately generating semantic and acoustic context and dynamically switching the generation process, it optimizes the overall song structure and sound quality.
-
Autoregressive Sketch Drawing: It uses an autoregressive model to generate music sketches, ensuring structural coherence and phoneme alignment.
-
Diffusion Model Refinement: It combines a diffusion model to perform high-fidelity refinement on the generated sketches, improving audio quality.
-
Combination of Discrete and Continuous Outputs: It uses discrete sketch tokens and VAE latents to output the final result, balancing structure and sound quality.
-
Multimodal Input Fusion: The input includes lyrics and audio samples, and the model achieves precise generation through multimodal fusion.
SongBloom's Project Address
- Github Repository：https://github.com/tencent-ailab/SongBloom
- HuggingFace Model Library：https://huggingface.co/CypressYang/SongBloom
- arXiv Technical Paper：https://arxiv.org/pdf/2506.07634
- Online Experience Demo：https://cypress-yang.github.io/SongBloom_demo/
SongBloom's Application Scenarios
-
Music Creation: It provides inspiration for musicians and creators, quickly generating high-quality song foundation frameworks, helping them explore new musical styles and creative directions.
-
Audio Production: In audio production for industries such as film and television, games, and advertising, it is used to quickly generate background music or theme songs, improving production efficiency.
-
Education Field: As a music education tool, it helps students understand music structure and the creative process, stimulating interest in learning.
-
Entertainment Industry: On platforms such as social media and short videos, it generates personalized music content for users, enhancing interactivity and fun.
-
Commercial Applications: Generate customized music for enterprises and brands, for use in product promotion, event publicity, etc., to enhance brand influence.