AIToolScan

YuE2 — Open Full-Song Music Generation, Compared to Suno for YouTube Creators

YuE is an open-source foundation model for lyrics-to-song generation: give it lyrics plus a style prompt and it writes a complete vocal song with accompaniment. The YuE2 release (technical report September 2026) adds symbolic planning — it first writes an editable melody-and-chord score in ABC notation, then realizes that score as audio. That score is the key differentiator for creators: you can read the composition, change the hook, and re-render before anyone hears it.

For a YouTube creator the pitch is not “free music” — it is licensing and Content-ID control. Suno makes great songs but its output is uploaded by thousands of other people to distributors, and those uploads land Content-ID claims on your videos. YuE2 renders locally on your own GPU, so nothing is fingerprinted by anyone else, and the model license explicitly lets content creators monetize outputs with no fees and no royalties. The trade-off is a 24 GB VRAM GPU and your own time.

License — the single most important row for creators

Read that carefully: an individual creator or musician is inside the free monetization permission. A studio, agency or label shipping YuE2 audio commercially is not — that needs a negotiated license. YuE also asks you not to use it for illegal, harmful, deceptive or unethical purposes.

Cost — what it actually runs you

Suno credits do not roll over between months, and only paid plans can buy extra credits or extra downloads. A creator publishing one polished track per week hits Pro’s 20-download ceiling fast — which is exactly the workload where a local model costs nothing.

Quality — WildSongBench, 192 prompts (September 12, 2026)

How to read this honestly: YuE2 only beats Suno when you sample 8 candidates and keep the best. A single YuE2 render (6.7316) sits between Suno v5 (6.8721) and Suno v5.5 (6.7150). The authors state plainly that the gap between the top means is not statistically significant. Suno still leads on MuLan (vocal-lyric alignment) and on PER (phoneme error rate — how cleanly syllables land on notes), which matters if you sing in Korean or Japanese and care about intelligibility. The honest summary: YuE2 matches frontier Suno on average musicality, wins on control, and loses on phoneme precision and turnkey speed. Note the original YuE 1.0 scores far lower (4.9165) — if a tutorial tells you to run “YuE”, check whether it is the v1 branch.

What you can do that Suno cannot

• White-box editing. Export plan/score.abc, have an agent or a text editor change harmony, melody, tempo or song form, and re-render. Suno edits audio, not composition.
• Zero-shot covers. Transcribe any recording with SheetSage2, review the melody ABC, then re-perform it in a new style with new lyrics. On 948 works, full-score covers hit 0.647 CLEWS mAP vs 0.006 with no score — the melody track is doing all the work, with no cover-specific fine-tuning.
• Reproducible seeds and artifacts. Every run saves its score, semantic tokens, acoustic latents, settings and model IDs next to the FLAC. A Suno render is a one-way door; a YuE2 render is a branch you can re-enter.
• Instrumental mode. Drop the vocal track out of the score and render BGM-only — the common YouTube case.
• No queue, no cap. Render overnight while you edit the next video.

What you give up: Suno Studio’s MIDI, stem separation, personas, custom voice, 8–30 minute audio upload and instant browser iteration. YuE2 also re-renders the whole recording after a score edit — there is no non-destructive stem-level edit.

Hardware — the real gate

• Linux, Python 3.12, NVIDIA GPU with BF16 support and 24 GB VRAM (RTX 3090 / 4090 class). Output is 48 kHz stereo with no quantization.
• On a consumer card, run one song at a time and expect minutes per render. That is fine for a 10-minute podcast bed, and slow for scoring ten hook variations.
• No GPU? The NOIZ-hosted demo is free in-browser, and ComfyUI ships native YuE2 nodes for people already running a local pipeline.

Quick start

git clone https://github.com/multimodal-art-projection/YuE.git
cd YuE
python3.12 -m venv .venv && source .venv/bin/activate
python -m pip install --upgrade pip && python -m pip install .
python examples/generate.py --output outputs/first-song

Settings: cot="full" (default — writes an editable melody+chord plan), cot="melody" (melody plan, free accompaniment — use for covers), cot="off" (straight from lyrics and style). Pass abc=... to supply your own score.

YouTube monetization checklist

• Disclose. YouTube requires creators to declare realistic synthetic or meaningfully altered content in Studio. AI music is not banned, but undisclosed AI content risks demonetization and, if repeated, suspension of monetization.
• Content ID is the bigger risk with Suno. Because every Suno user can distribute the same output, third-party uploads generate claims on your video. A claim is not a strike and is not a channel-termination risk, but it diverts or blocks revenue until you dispute it — and other users report Suno-made tracks being taken down by distributors and platforms. Locally rendered YuE2 audio carries no such registry entry.
• Copyrightability. Purely AI-generated audio is generally not copyrightable on its own; protection attaches to human authorship. The YuE2 score workflow is the practical answer — you author and revise an actual composition document, which is the human-created artifact worth registering.
• Distribution is unsettled. DistroKid, the distributor most AI-music creators use, is currently being sued by Universal over AI-generated uploads. Keep distribution optional and reversible; a self-hosted YouTube channel does not depend on it.
• Keep a prompt log. Save lyrics, style prompts, seed and the ABC score with the project. song.save_artifacts() does this automatically and it is your provenance record.

Honest limits

• The 24 GB VRAM requirement excludes a lot of creator laptops, and there is no hosted fallback other than third-party demos.
• Best-of-8 selection is what produces the headline number — budget for it, or you are at single-render quality.
• Phoneme error rate trails Suno v4.5/v5.5, so heavily syllabic Korean or Japanese vocals may need more candidate passes.
• Company-scale commercial use of the weights is not covered by the free permission.
• Length caps at roughly five minutes per generation.

Links

GitHub: https://github.com/multimodal-art-projection/YuE · YuE2-3B weights: https://huggingface.co/m-a-p/YuE2-3B · Technical report (Sept 2026) linked from the repo · Demo page: https://map-yue.github.io/ · Original YuE paper: arXiv 2503.08638 · WildSongBench prompts and results for reproduction: https://huggingface.co/datasets/m-a-p/WildSongBench