Our indie band has been testing Suno's new extended generation mode for a few weeks. We primarily use it for drafting song structures and generating instrumental sections to build on.
The 3+ minute tracks are definitely a step up for creating more complete demos. The coherence over the longer duration is impressive, but we've noticed the "vibe" sometimes shifts unexpectedly around the two-minute mark, like it forgets the initial guitar tone. Has anyone else run into this, or found a prompt technique to keep it more consistent? We're still figuring out if the output is detailed enough to justify using it as a core part of our workflow, or if it's better for initial inspiration.
Totally see what you mean about the vibe shift. We've had the same issue with the guitar tone changing mid-track. It's like the model starts a new "section" and resets some parameters.
What's worked for us, somewhat, is being super repetitive in the prompt. Instead of just "indie rock with clean arpeggiated guitar," we'll literally write "maintain the same clean, chiming electric guitar tone with medium reverb throughout the entire track." We'll mention it two or three times. It's not perfect, but it seems to help anchor it a bit.
For our workflow, it's firmly in the "spark" category. The output isn't detailed or consistent enough to be a core piece, but it's fantastic for getting past the blank page. Have you tried feeding it a short clip of your own guitar tone as a reference? I've heard that can help but haven't tested it yet.
Benchmark or bust
Yeah, the coherence drop around two minutes is a real thing. We've seen similar gaps in logic when generating longer synthetic log streams for testing. The model seems to hit a context window boundary and soft-resets.
For tonal consistency, you might try "seeding" it. Instead of just a text prompt, could you feed it a two-bar loop of the exact guitar tone you're after as an audio input? Even a rough phone recording. That sometimes gives the model a stronger anchor than descriptive text alone.
It sounds like you're already using it the right way, as a sketchpad. I'd be skeptical about integrating it directly into a core workflow without a human to catch and correct those vibe shifts. It's a fantastic muse, but not a reliable bandmate yet.
—jr