YuE2 in ComfyUI: A 3:16 Indie Punk Song in 95 Seconds on a 4080

YuE2 is the new native ComfyUI text-to-music stack: one style line, tagged lyrics, a complete song with vocals and band. I wrote an indie-punk lyric with Grok, dropped it into the official YuE2: Text to Music template, and got a 3:16 stereo MP3 in 95.01 s on an RTX 4080. The listening quality is the first local vocal song on this box that I would actually keep.

What makes YuE2 different from MiniMax Music 3 / ACE-Step is the optional ABC planning pass: the model writes a symbolic melody (and chords in full mode) before it renders audio. Docs: ComfyUI YuE2 guide · weights: Comfy-Org/YuE2 · upstream: m-a-p/YuE2-3B.

What I ran

GPU: NVIDIA GeForce RTX 4080 (16376 MiB), driver 610.88
ComfyUI: 0.36.0 (YuE2 needs 0.35.0+; 0.35.1+ recommended on AMD)
Template: Audio → YuE2: Text to Music
Node: Text to Music (YuE2) → Save Audio (Advanced) (MP3, quality V0, prefix audio/YuE2)
max_duration: 240 s · mode: full · ABC_planning: true · seed: 6
Output: output/audio/YuE2_00011.mp3 — 196.56 s (3:16), 48 kHz stereo, ~240 kbps, 5.6 MB
Wall clock: 95.01 s (ComfyUI output tile)

Faster than real-time for a full vocal track: ~197 s of audio in 95 s of GPU time. The song stopped short of the 240 s cap, which matches the template note — duration is a ceiling, not a pad.

Template card

Files and folders (this machine)

One checkpoint. That is the whole install for text-to-music:

• models/checkpoints/yue2_3b_int8_convrot.safetensors — 3.69 GB (3960938800 bytes)

CheckpointLoaderSimple inside the subgraph pulls MODEL + CLIP + VAE from that file. I did not have sheetsage2_bf16.safetensors on disk — that audio encoder is only for the YuE2: Music Cover template, not this run.

Graph

The canvas is two nodes. Open the subgraph and the real path is:

• CheckpointLoaderSimple → YuE2GenerateABC (if ABC planning is on) → YuE2GenerateMusic
• seconds out of Generate Music feed EmptyYuE2LatentAudio
• KSampler (template defaults: 32 steps, CFG 1, dpm_2 / sgm_uniform, denoise 1) → VAEDecodeAudio

Leave those sampler values alone unless you are debugging. The pins that matter on the subgraph are Style, Lyrics, mode, ABC_planning, max_duration, seed, ckpt_name.

Settings that matter

• Style — genre, vocal, language, tempo, instruments. Keep musical direction here, not in the lyric box.
• Lyrics — only words meant to be sung, plus section tags: [Verse], [Chorus], [Bridge], [Outro]. I also used [Pre-chorus] and it followed the form.
• mode — full = melody + chords (this test). melody = melody only, better for covers.
• ABC_planning — true for a symbolic plan first (recommended). Off skips ABC and generates from Style + Lyrics directly.
• max_duration — ceiling in seconds, up to 900. Long lyrics get an automatic token-budget trim so they still fit context.

Style and lyrics I used

I drafted the lyric in Grok, then pasted it into the node. The chat asked for a sexy female vocal; the ComfyUI run used a male raw voice in Style. Report the graph, not the chat.

Style

English, fast awesome indie punk rock , male raw voice, guitar riffs and love

Lyrics (as pasted into the node):

[Intro – dirty guitar riff, feedback, then it kicks]
[Verse 1]
Last call lights and cheap perfume
You wipe the bar like you own the room
Blonde hair falling in your eyes
Leather jacket, no disguise
I come here every goddamn night
Just to watch you pour the fight
Out of other people’s mouths
While I’m drowning in your mouth

[Pre-chorus]
Don’t look at me like that, Mara
You already know

[Chorus]
Mara, Mara, wreck me slow
Keep that pint glass in your hand
Sexy blonde who runs this show
I’d burn the whole pub down for you
Guitar screaming through the floor
Heart on fire, nothing left to lose
Mara, I’m in love, it’s true
Come on, baby, make it bruise

[Verse 2]
You lean across and steal my smoke
Laugh like you already know the joke
Every guy in here’s a ghost
I’m the only one who stays the most
You got that look that starts a riot
Then you act like you don’t buy it
I want your name on my neck
And your lipstick on my wreck

[Pre-chorus]
Don’t look at me like that, Mara
You already know

[Chorus]
Mara, Mara, wreck me slow
Keep that pint glass in your hand
Sexy blonde who runs this show
I’d burn the whole pub down for you
Guitar screaming through the floor
Heart on fire, nothing left to lose
Mara, I’m in love, it’s true
Come on, baby, make it bruise
Bridge (half-time, then it explodes)
I don’t want last call
I want the keys
I don’t want last call
I want your teeth
Final chorus (bigger, shouted, riff going feral)
Mara, Mara, wreck me slow
Keep that pint glass in your hand
Sexy blonde who runs this show
I’d burn the whole pub down for you
Guitar screaming through the floor
Heart on fire, nothing left to lose
Mara, I’m in love, it’s true
Come on, baby—
make it bruise

[Outro]
(riff repeats, voice getting closer, almost a whisper then a snarl)
Mara…
one more round.

Official rule: put only sung words in Lyrics. This paste still has intro/bridge/outro stage directions mixed in, and the song still held. Next take I would move those lines into Style and keep [Bridge] / [Outro] as tags only.

Hardware test

Seed 6, INT8 ConvRot 3B checkpoint, ABC on, mode full, 240 s cap. Result is a 48 kHz stereo V0 MP3 at 3:16. Vocals are intelligible, the guitar riff actually arrives, and the chorus comes back as a chorus — that is the bar MiniMax Music 3 already cleared, and YuE2 cleared it on a single 3.69 GB file instead of a DiT + encoder + VAE trio.

Checklist

1. Update ComfyUI to 0.35.0+ (this box is 0.36.0).
2. Template Library → YuE2: Text to Music → download yue2_3b_int8_convrot.safetensors into models/checkpoints/.
3. Write Style (language + genre + vocal + instruments). Write Lyrics with [Verse] / [Chorus] tags only.
4. mode full, ABC_planning on, max_duration 120–240 while you learn the voice, then go longer.
5. Save Audio (Advanced) → MP3 V0. Iterate seed / Style with the same lyric.

Bottom line

I am keeping YuE2 installed next to MiniMax Music 3. For a tagged lyric + a short style line, this native node is the faster daily path on 16 GB: one checkpoint, ABC you can inspect if you open the subgraph, and a 3:16 punk song that sounded finished in a minute and a half. Music Cover (SheetSage2 transcription) is the next template to try; that extra encoder is not required for text-to-music.