DiffRhythm AI turns two simple inputs — lyrics and a style prompt — into a finished song. The underlying model is described as the first latent-diffusion-based song generation approach capable of synthesising complete songs with vocals and accompaniment together, for up to about four minutes in the base model and longer on paid tiers. Because generation is non-autoregressive, output arrives quickly, and there is no need to generate vocals and backing tracks separately through a multi-stage pipeline.
The workflow is deliberately minimal: paste your lyrics, describe the genre or mood, hit generate, then download. Style prompts can be as simple as "pop", "rock", "jazz" or a longer descriptive sentence about instrumentation and tempo. The site also publishes a public gallery of community-generated tracks, which doubles as a way to see what the model handles well across genres.
Important context: this website is an independent third-party platform. The DiffRhythm research itself comes from the ASLP Lab, and DiffRhythm.ai states clearly that it is not affiliated with that lab. Users should also check the licensing terms of their chosen plan before using generated audio in commercial projects, since commercial rights are tied to paid subscriptions.
