Inside the YuE2 Browser Studio
Upon visiting YuE2 at yue2ai.app, I found a page that looks less like a consumer music toy and more like a stripped-down DAW front end. The composer sits above the fold: three tabs — Lyrics to Music, Edit Music and Instrumental — a prompt field, a title box, a lyrics editor with an "@" shortcut for inserting section tags like Verse and Chorus, a Style field with 1160 characters of headroom, and a row of presets (indie pop, jazz and soul, electronic, acoustic).
A credit counter above the Generate button reads "3 credits this generation," and the advanced panel exposes a seed field with the note that -1 returns a random result and that reusing a seed lets you compare variations. There is also a separate "AI write / edit lyrics" button that costs 1 credit per request and holds drafts in a review pane so that "your song lyrics stay unchanged until you apply a draft." That small detail — a draft you must explicitly accept — is the kind of thing I rarely see in this category.
In testing the free tier, I signed in with the 5 starting credits and ran a single Text to Music generation: lyrics divided into Verse and Chorus, a style line describing warm indie pop with acoustic guitar and gentle drums, seed left blank. The progress indicator climbed steadily toward 95% and then sat there, exactly as the documentation warns. The finished file played inline and downloaded without a paywall — a small but telling detail, since several competitors watermark or lock downloads.
How the Symbolic Planning Step Actually Works
What separates YuE2 from most lyric-to-song services is that the model composes before it performs. YuE2-3B, released by M·A·P (Multimodal Art Projection), runs a stage called symbolic planning: it writes melody and chord progression in ABC notation, then renders audio from that plan. In the official model workflow you can open the score between those two stages, change a chord, and render another version.
The official API exposes three planning modes — melody-and-chord planning, melody-only, and generation with no symbolic plan at all. The browser composer on this site deliberately does not surface that selector; it also does not offer note-level editing or score export. So you get the model's audio quality with a friendlier interface, but you lose the very thing the model is famous for. For score inspection you need to run the model yourself: the local quick start specifies Linux, Python 3.10+ and a 24 GB NVIDIA GPU with BF16 support, with community ComfyUI workflows available separately.
The studio does give you a partial workaround — an Audio to Sheet Music tool that converts a recording into downloadable ABC notation and a score image, plus Audio to MIDI for a .mid file you can drop into a DAW. In testing, I would describe this pairing as "generate here, analyse there," rather than a single editing environment.
Pricing: Credits, One-Time Packs, and Commercial Rights
The free tier gives you 5 credits — enough for exactly one Text to Music generation at 3 credits, with 2 credits left over. Text to Music runs 3 credits; Music to Music, which accepts a reference recording, runs 4. Lyrics writing and editing is 1 credit per request. Paid packs are one-time purchases, not subscriptions: Starter is $9.90 for 99 credits (about 33 songs), Basic is $29.90 for 370 credits (about 123 songs), and Plus is $49.90 for 700 credits (about 233 songs, 10 concurrent tasks). All tiers include unlimited downloads, and paid plans add unrestricted commercial use with no royalty or revenue share.
Per-credit cost falls from roughly $0.10 on Starter to about $0.07 on Plus, with no monthly commitment anywhere on the pricing page. Compared with subscription-only rivals that gate commercial rights behind a monthly fee, the one-time model is genuinely attractive for hobbyists who generate in bursts.
Where YuE2 Stands Against Suno v6
The team published a September 12, 2026 comparison across 192 WildSongBench prompts. Standard YuE2 scored 5.9075 on musicality versus 5.6558 for Suno v6, and 6.7316 on SongBench average versus 6.5562. Suno v6 won on pronunciation: 8.44% phoneme error rate for YuE2 against 7.58%, and higher AllMusicCaps text alignment at 0.4305 versus 0.4054. Read this as a model-level evaluation, not a promise about your next generation.
Against Suno more broadly, the trade-off is control versus polish. Suno's editing tools and vocal clarity are more forgiving out of the box; YuE2's appeal is that the composition is visible and inspectable, and that the weights are downloadable for researchers. If you want a finished-sounding track in two clicks, Suno remains easier.
Genuine Strengths and Real Limitations
Strengths first: the open-weight release with published evaluation material is unusual in this space, the ABC-to-audio pipeline makes the model useful as a research object, and the studio's unlimited downloads plus royalty-free commercial licence on paid packs is a fair deal. The public demo gallery — six official examples spanning English, Chinese, Japanese and Spanish — shows wide stylistic range, from cyber metal to flamenco.
Balanced against that, four limitations surfaced in my testing:
- The studio hides the planning-mode selector and offers no score export or note-level editing.
- Pronunciation is the weak point; the 8.44% phoneme error rate means you should audition every vocal line, particularly on non-English lyrics.
- The progress bar rises to 95% in about 60 seconds and then waits — it is a visual estimate, not a measurement.
- The editing agent described in the official demonstrations is not included here.
The interface itself is plain but fast: no onboarding tour, no template gallery, just fields and a Generate button. For a tool with this much model complexity underneath, that restraint works in its favour — though the absence of a score viewer in the browser will frustrate exactly the users the marketing targets.
Who Should Use YuE2 — and Who Should Look Elsewhere
YuE2 suits songwriters who already have finished lyrics and want a vocal demo to judge phrasing against, producers curious about inspecting melody and harmony as data, and developers evaluating open music models. It is also one of the few services where a $9.90 one-time purchase buys commercial rights with no strings attached. Several gallery tracks show play counts in the hundreds of thousands, suggesting an audience beyond launch-day curiosity.
Look elsewhere if you need broadcast-ready vocals with minimal cleanup, or if you specifically want the score-editing workflow — in that case run the weights locally on a 24 GB GPU. For everyone in between, the honest recommendation is this: start with the 5 free credits, generate one song from your own lyrics, listen closely to the consonants, and decide from that single take whether the control YuE2 offers is worth the extra review work. If it is, the Plus pack is the better value per song.
Visit YuE2 at https://yue2ai.app to explore it yourself.
Comments