Local AI Music You Can Ship: Why We Run ACE-Step 1.5, Not YuE2

The model that claims to beat Suno

YuE2-3B arrived on Hugging Face in October. Its card puts it at the top of WildSongBench, a 192-prompt song benchmark: a SongBench average of 6.9632 against 6.8721 for Suno v5 and 6.5562 for Suno v6. The authors report a 3.6-minute song in 71 seconds on an RTX 4090 at 11.18 GiB of VRAM, which would fit our 16 GB card.

Three things in that claim deserve a second read:

• Best-of-8. The headline number generates eight songs and keeps the best one. The standard setting picks the better of two and scores 6.7316, below Suno v5.
• Whose benchmark. WildSongBench comes from the same team. The blind listening arena they link to is still collecting votes.
• The licence. The weights are CC BY-NC 4.0: non-commercial use only.

Why the licence decides it

Everything we make ends up on a brand account or in client work: OMNI Dramas, the late-night show, ad films. That is commercial use in the plain sense of the licence, whether a given post earns money or not.

We also asked the obvious follow-up: what if we generate with YuE2 and then rework the track? Our reading is that it does not help. The licence covers copying and running the weights, so the use happens when the song is generated, not when it is published. Reworking the output adds human authorship to the parts you changed; it does not change how the rest was made. We are a studio, not a law firm, so treat this as our working rule, not legal advice.

The same rule already applies to an image model in our stack, Qwen-Image 2.1, whose research licence keeps it on spec work. A music model under a non-commercial licence would sit in the same box, and we have no spec work that needs music badly enough to justify the install.

The commercial-clean alternative

ACE-Step 1.5, from ACE Studio and StepFun, is in the same table at 6.0118. Lower, but the licence is MIT, and the model card says plainly that generated music can be used for commercial purposes. It also says what the model was trained on: licensed tracks, royalty-free and public-domain music, and synthetic audio rendered from MIDI.

The gap in the benchmark is real. We decided a model we can ship from is worth more than a better one we cannot, and then checked whether ACE-Step was good enough by listening.

How we ran it

• One RTX 5060 Ti with 16 GB, the same card as everything else on this Journal.
• ComfyUI 0.42, which ships ACE-Step 1.5 nodes and ready-made workflows. Nothing had to be updated or installed besides the weights, about 28 GB in total.
• Two presets from the stock workflows: turbo (2B diffusion model with 0.6B and 1.7B Qwen text encoders) and XL (4B diffusion model with 0.6B and 4B encoders).
• 8 steps, cfg 1, euler sampler, simple scheduler, shift 3, with the language-model planner on. That planner writes the song's structure before the audio is generated.
• Output: 48 kHz stereo FLAC.

Three test briefs: an 80s synth-pop song with a female vocal and lyrics we wrote for the test, a suspense underscore for a drama, and a 1970s talk-show house band.

The numbers

Both presets run many times faster than the music plays. That changes how you work: you don't rewrite the prompt until a take comes out right, you generate several seeds and pick by ear.

Listen

The same lyrics and seed through both presets. Each player is the raw track, loudness-normalised and nothing else.

Two instrumentals of the kind we actually need for a drama and for the show:

The covers were generated on the same card with MOSS-Image, the local model we use for images with readable text. Each took 9 to 15 seconds. The band cover needed a seed sweep: the title was spelled right on 2 of 4 seeds, which is the usual rate for that model.

Did it sing the words?

We can't measure taste, but we can check the lyrics. We ran both versions of the song through Whisper (the small model, on the CPU) and compared the transcript with the lyric sheet.

• Turbo sang the first verse word for word. Later sections came back only in parts.
• XL sang the pre-chorus, both choruses and the whole second verse almost word for word. Whisper found no words in its first 20 seconds, so either the intro runs long or the first verse was skipped; only listening tells which.

This is one seed per model, so it is a single comparison, not a ranking. Whisper also mishears sung vocals over a full mix: it heard "Neon static" as "Knee instead" more than once. Turn off its voice-activity filter for music, or it treats the singing as noise and returns nothing at all.

On quality, the final call was a listen: all five test tracks were good enough to use. We now start vocal tracks on XL and keep turbo for drafts and instrumentals.

Before it goes into a mix

The vision models we use to check video can't hear, so audio QA here is three checks: Whisper for the words, a loudness reading for the level, and a person for everything else.

Where this leaves our pipeline

YuE2 stays bookmarked. The first YuE is Apache-2.0; if YuE2 ever ships under a commercial licence too, its editable scores and cover mode would be worth a test.