YuE2: local music generation, GPU setup and costs

YuE2 produces music rather than spoken TTS. Browse speech and audio models for related workflows; the TTS calculator compares character-billed speech APIs, so it does not price a generated song.

For a new song, full planning produces a melody and chord score before audio synthesis. Melody-only planning leaves the accompaniment more freedom, while turning planning off generates directly from lyrics and style. You can also supply an ABC score.

For covers, the documented workflow uses SheetSage2 to transcribe the source recording, then feeds its melody score and a new style or lyrics into YuE2. Editing a score renders a new recording.

The current repository recommends Linux, Python 3.12 and an NVIDIA GPU with BF16 support and 24GB VRAM, starting with one request at a time. It produces 48kHz stereo without quantization. Model files download from Hugging Face on first use.

Follow the current source install below. The Hugging Face card also documents an older wheel install with different Python requirements; use the instructions for the package version you actually install. Save outputs in a fresh directory so timings and failed attempts remain available.

Open outputs/first-song/audio.flac, then inspect the saved score, configuration and generation record. save_artifacts() retains the audio, symbolic plan when used, semantic tokens, latents, timings and model identities. Check the truncated flags: a playable audio file can still have reached a token limit.

To revise harmony or melody, copy the score into a new ABC file and submit it as a new input. Pin both model and VAE revisions for comparisons. The default YuE2-Vae is for generation and listening; YuE2-Vae-legacy is the decoder used in the reported benchmark protocol.

Downloading the weights has no per-song API charge from the authors. Local costs are the GPU or full cloud instance, billed runtime, storage and the time spent selecting or editing outputs. Measure elapsed generation time for your song length and settings before estimating a monthly budget.

If an instance costs H dollars per billed hour and the complete session takes T hours, its compute charge is H × T. Include setup, idle time and repeated candidates in T. Budget two candidates for the standard benchmark setting and eight for best-of-8.

The code and documentation use Apache 2.0, while YuE2-3B and the VAE weights use CC BY-NC 4.0 with additional individual creator permission. Personal users, creators and musicians acting individually can monetize generated outputs without model-license fees or royalties to the authors, subject to the additional terms. Companies need to discuss a separate commercial weight license; that price is not published. Arrange rights to the source song when making covers.

Lyria 3.5 is a hosted music-generation alternative through the Gemini API. Google lists $0.08 per full song in the paid tier, with no free-tier API offer. Billing is per song/request. At that listed rate, 100 songs cost $8 and 1,000 cost $80, before other applicable charges.

Google’s guide describes text or image prompts and 44.1kHz stereo output, with full songs lasting a couple of minutes and duration influenced by the prompt. YuE2 provides local weights and an editable symbolic score; choose it for score control, or Lyria for a hosted workflow.

The authors’ WildSongBench comparison covers 192 prompts. Both reported YuE2 settings use symbolic planning and the legacy evaluation decoder; standard candidate selection and best-of-8 have different compute costs.

Listen for lyric pronunciation, musical structure, vocal quality and whether the intended edits took effect. Retain the input, score, revisions and failed attempts when comparing candidates. For spoken narration rather than a song, HeyGen Voice is a different workflow built around cloning and speech synthesis.